I recorded a two-minute video about one AI technique I’ve been using. I wrote about it previously in The Writer’s Room. It’s a Claude Skill I built called the Cast of Critics, and it’s the part of my writing process I get asked about most. I figured I’d show up on camera and talk through it.

I’m curious whether anyone else is using AI in a similar way. For critique, structured review, anything that helps you stress-test your thinking before you commit to it. Doesn’t have to be writing. Could be how you review a product decision, a pitch, a strategy doc.

If something resonates here, or if you think I’m doing it completely wrong, click the button below to send me a message.

More long-form coming soon. This is me trying something new.

When Paramount/SkyDance announced earlier this year that they would shut down MTV’s music channels after already shutting MTV News and its archives, there were a lot of wistful elegies for the culture we once had. FX has built a cottage industry of dramatized series set in the era. You hear how there is no more independent film or independent music. Media ownership has consolidated into fewer and fewer hands — studios merging, private equity rolling up what’s left of local news, billionaires acquiring the infrastructure of public discourse.1

These things are true.

It also misses the point. Culture isn’t gone. It’s just not found in traditional media spaces if that’s where you were looking.

A filmmaker with something to say today isn’t waiting for Sundance to validate them. They’re on YouTube before they’d ever approach a distribution system that would probably lock them out anyway. A musician shaping culture right now isn’t coming from an indie rock scene most people lamenting the collapse are still listening to. They’re often found on TikTok or SoundCloud — and more likely to come from culture originating outside the U.S., like Reggaeton, K-Pop, Afrobeats, genres that didn’t need the American music industry’s permission to reach American audiences.

Distribution was a chokepoint in the 90s. Something else took its place.

Distribution Was the Chokepoint. The Internet Fixed It.

In the 90s, independent culture was vying for access to a system it couldn’t fully control. Sundance was a door to theatrical distribution. Sub Pop and Matador were a door to record store shelves and radio play. Image Comics — founded in 1992 when seven of Marvel’s top artists walked out and started their own creator-owned studio — was a door to comic shop distribution and mainstream retail.

The cultural energy was real. Grunge, hip-hop, independent film, creator-owned comics. People making things outside the institutional system, and audiences hungry for them.

But they were still largely dependent on that legacy distribution system to reach audiences. Getting noticed required getting into the system — or building enough pressure outside it that the system came to you.

Over the next 30 years, the internet collapsed the cost of distribution to near zero. The pipe opened to everyone. I wrote about what this meant for creators in a recent piece, “You’re In the Creator Economy Now.”

Now there’s a new chokepoint, and that’s where this piece picks up.

The Algorithm Is the New Gatekeeper — and It’s Invisible by Design

The 90s gatekeeper was visible. If you were a filmmaker and your film couldn’t get into theaters, that’s because it didn’t have distribution. If you were a musician and your music wasn’t in stores or played on the radio, it’s because no record label picked it up. The barrier was a known and named system. You could see the door and figure out how to knock on it — or create pressure through cultural resistance. Most of the cultural breakthroughs of the 90s happened this way.

Today, the chokepoint is the social media algorithm. It’s a ranking system with rules and mathematical probabilities encoded by computers to match content to individuals, tuned and managed by engineers at the direction of a small handful of companies to maximize their revenue. It is largely invisible and mysterious to the creators and audiences it mediates.

— researcher and social media scholar who has studied these platforms for two decades — recently named what happened. She argues that social media has become parasocial media: platforms built not for genuine connection between people, but for one-directional performance optimized for engagement metrics.

Audiences don’t experience this as withholding. They experience it as an endlessly scrollable feed — a bottomless recommendations tab of content “you might also like.” They don’t see what’s being filtered out. They only see what the platform decided to show them, optimized for the engagement metric that serves the platform’s business model.

This phenomenon doesn’t only affect creators seeking an audience. It affects anyone navigating these platforms with a purpose — including job seekers. I feel it on LinkedIn every day. My network doesn’t see what I publish. The algorithm decides who gets it, when, at what volume — and the logic is opaque by design.

The distribution chokepoint moved from physical retail and broadcast infrastructure to platform algorithms. The gatekeeping function didn’t disappear. It became invisible and unaccountable. And it affects both sides: creators can’t reliably reach audiences they’ve built, and audiences can’t reliably find things they’d want if they could find them. The platform captures the value that should flow between them.

About a decade ago, there was a symbiotic relationship between platforms and publishers — platforms sent traffic to publishers. That era is over. The chokepoint is in effect. Social media traffic is gone. Search traffic is rapidly declining. So as a publisher or creator, how can you resist and connect with your audience?

Substack Isn’t a Safe Harbor. It’s a Better Bet.

The conventional response to platform dependency is to build a direct audience. Email list. Owned channel. Something the algorithm can’t touch. That’s the actual asset — not the follower count as mediated by a platform, but a verified list of people you can reach who are opted in.

That’s what I’m trying to do here on Substack. And if you’ve spent any time here, you know I’m not alone.

But why Substack specifically, rather than any email tool? Isn’t this just another platform that will end up the same way?

Three reasons, in order of importance.

First: incentive alignment. Substack makes money when creators make money — through paid subscriptions, not advertising. That’s a structurally different relationship than a platform selling the audience’s attention to advertisers. LinkedIn’s incentive is to keep you scrolling, consuming, and increasingly interacting with ads — or content masquerading as ads. Substack’s incentive is to help you convert readers into paying subscribers. Those produce different product decisions.

Second: network effects that other newsletter tools don’t have. Notes — Substack’s short-form feed, where writers share observations and restack each other’s work — create real social mechanics for discovery within the platform. It is still early stage and more organic than what LinkedIn or Twitter eventually became, but the same building blocks. Which means Substack has a plausible path to becoming a meaningful discovery layer for written work.

It also means Substack is subject to the same enshittification trajectory identified — and that danah boyd has applied to social platforms more broadly. A platform that starts aligned with creators can drift toward optimizing for its own engagement metrics as it scales. Substack is not immune. I’m watching it with eyes open.

Third: email-native delivery. A subscriber who signs up gets my writing in their inbox regardless of what Substack’s algorithm does that week.

That’s the bet. Not that Substack is permanent. That its incentive structure, network mechanics, and email-native delivery make it a better foundation than the alternatives available right now.

The Agent That Replaced My Feed

Deciding to build on Substack told me where I wanted the audience to live. It didn’t tell me how to find the people worth engaging with to grow it.

The answer to that problem, on any platform, is supposed to be the feed. Browse, scroll, discover, engage. But the feed is the algorithm’s answer, not mine. What surfaces there is optimized for platform engagement, not for my specific interests or the specific neighborhood of writers I’m trying to be part of.

So over spring break I built an alternative. I’m calling it the Signal Pipeline.

Every morning it scans the 44 Substack publications I identified through a combination of my own seed list and a crawl of the recommendation graphs of writers I already trusted — all working at the intersection of media, product leadership, AI, and the creator economy. For each new post, it runs a relevance score against five topic clusters derived from my own reaction history: what I’ve actually engaged with across two years of reading and reacting on Substack and LinkedIn. The scoring isn’t guessing at my interests. It’s derived from what I’ve already demonstrated I care about.

Posts that score above a threshold get a second pass. The system fetches the full text and uses an AI model to identify the most quotable passage and suggest two or three commentary angles specific to my background — my NBC News experience, my current positioning, the argument I’m already making in my writing. Posts that don’t clear the threshold still appear in the digest with their scores, so I can sanity-check what the system is skipping.

By the time I’m at my desk, the digest is already waiting. Review takes about five minutes. Then I choose what to engage with and write the commentary myself — typically ten to fifteen minutes per restack. What used to take one to two hours of manual browsing is now a five-minute review. The writing is still mine.

Before this: 11 articles published over six weeks, ending in March. 26 subscribers. Spring break, I built the pipeline. Mid-April, I started the restacking habit. Two weeks later: 42 subscribers. I’d published 11 articles before without this kind of movement. Content alone wasn’t moving the needle. Showing up consistently in the Notes feed, engaging with specific authors and their work — that’s what changed.

I’m not claiming causation. I’m reporting a before/after that’s hard to ignore.

What I’m doing is small. But I don’t think I’m the only one heading this direction.

Boyd named the disease. My LinkedIn feed is the symptom. What Dan Porter — serial entrepreneur, most recently founder and CEO of Overtime Sports — and — who helped define the Product Led Growth movement and is now Head of Growth at Lovable — are describing from their respective corners is the early shape of a response.

Porter recently wrote about agents replacing the open browser tab on the consumer side. You don’t browse for a restaurant or search and scroll fantasy football analysis on websites. You tell your agent what you want. The platforms and publishers currently sitting between you and that answer are the ones at risk.

Verna is making the same argument from the B2B side: if your product isn’t accessible to agents, you’re not part of the workflow anymore. A human user may never touch your product. The agent decides and does things.

Consumer behavior, B2B software, media audience development — three completely different problems, same structural observation.

You decide what signal looks like. You decide what’s worth engaging with. The algorithm doesn’t get to make that call anymore.

That’s worth building toward. Even at this stage.

Leave a comment

Summer 2000. My first day as a web design intern at a startup in Northern Virginia, a few miles from AOL’s campus.

Two businesses shared the same office: a services division making websites for clients, and a VC-fueled dotcom chasing music and e-commerce audiences. I was on the services side, hand-coding HTML for politicians and country musicians launching their first websites. The web was so new that I didn’t know if “making websites” could even be a career — it was just something I was interested in. I didn’t know that I would spend the next 25+ years in an industry that was still taking shape and wasn’t yet taken seriously.

I walked into the office on my first day — the dotcom bust had already hit. The startup folded. The office was empty. I never met a soul on that side of the building. Plush toys with the startup’s logo were littered around; nobody would even remember the company name. They sold all the Herman Miller chairs.

I didn’t panic. The jobs lost had barely existed a year. The opportunity still felt full of potential. That summer, I taught myself ActionScript in Macromedia Flash while listening to a new livestreaming music site called SoundBreak, another dotcom that folded within the year. I barely thought about it at the time.

That was the first time I watched a technological wave displace a generation of workers. But it didn’t feel like the end. The potential was boundless and the field was clearing.

I’ve been thinking about this moment a lot lately. Not because of the bust. Because of what always comes next.

A Clearing, Not a Collapse

On March 1st, I published a Substack piece about the structural changes hitting the knowledge worker labor market. President Trump had just touted low unemployment numbers in his State of the Union Address. But if you’re in the knowledge worker economy here’s what it looks like — the labor market is frozen, a tsunami of AI-cited layoffs are already hitting or rapidly approaching, and entry-level rungs have completely vanished.

The trend pieces caught up the following week. Now it’s impossible to miss what’s happening.

Knowledge work is being reshaped in real time. If you work in software, you’re feeling it most directly. While the dust is still settling, it’s already becoming clear that the way we’ve operated for decades is radically changing, and that means the shape of teams and organizations will change as well.

Two pieces I’ve been reading recently describe what this looks like from inside companies. Robonomics — written by a partner at Altimeter Capital — argues the real bottleneck in organizations was never speed, it was translation cost: PM to design to engineering to QA, every handoff losing fidelity, every translation requiring alignment meetings. AI collapses that. Craig Hepburn, a former CDO at Art Basel and UEFA, maps the five operational layers that replace the org chart when agents can run them. Both describe what the company gains from the redesign. Neither describes what happens to the people on the other side of it.

The dotcom bust looked catastrophic from inside the companies that collapsed. From the outside — from the position of the person who was just interested in the technology and had no company to protect — it looked different. The field cleared. What was left was cheaper and more accessible tools, faster infrastructure, and people genuinely curious about building what came next.

Most people who lost jobs in 2000 went back to work for companies, some stayed in the digital space and some didn’t. The ones we remember are the ones who stayed in digital and built the technologies that have come to define how we live, communicate, and collaborate today — search, e-commerce, streaming, social media, SaaS, cloud, mobile. All these technologies were still early or hadn’t yet arrived. Google was still a small private company when the dotcom bust hit. It didn’t IPO until 2004. Steve Jobs had only recently returned to Apple. They hadn’t yet launched the iPod. None of what came next was inevitable until it was.

Flash forward to now — it is literally the best time ever to build something if you make content or software. It’s radically easier and with fewer barriers than have ever existed before — technically, financially, operationally. That’s the part worth examining.

What’s Different Now

None of what follows is completely new. It’s been building for decades. Creators have been growing direct audience relationships since YouTube launched in 2005. Stripe has been processing payments for small businesses since 2011. The tools for building and shipping software have been getting cheaper and more accessible. The path from idea to paying customer was already there.

What changed is the friction.

The skill floor dropped. The time cost collapsed. Things that used to require months of learning, a technical co-founder, or a production budget you couldn’t afford — you can now start on your own, this week, without asking permission or raising money.

Three things in particular that fuel the digital economy:

Content creation and distribution. For most of media history, reaching an audience required a broadcaster’s or publisher’s backing: a network, a masthead, a platform deal, a distribution agreement. That business structure decided what counts as “professional” and who got to call themselves one. The gatekeeper wasn’t just a filter; it was the only path to the audience.

That’s no longer true. A writer, a video creator, a domain expert with something worth saying can now build a direct relationship with an audience and get paid for it — without a gatekeeper deciding whether they’re worth the investment. The tools for creating, distributing, and monetizing content are available to anyone. What’s required is the discipline to show up consistently and the judgment to make something worth paying attention to.

Building software. Software development used to require a team with distinct specialized skills — a product manager to define the work, a project manager to coordinate it, engineers with further specialized skills to build it, designers to shape the experience, QA to test it and catch what broke. That team took weeks or months to ship anything meaningful. What previously required all of that can now be done almost entirely by a single motivated person with a clear idea and the right AI tools, in days. I’ve built data pipelines, editorial workflows, and content systems over the last few months — projects that would have required a developer, a data engineer, and significant runway in any prior period. A few months ago this was inconceivable. Now it’s just where things are. Building a digital product right now is like snapping together LEGOs. The pieces exist. Assembling them doesn’t even feel like work.

Making money on the internet. The infrastructure for getting paid used to be a serious barrier. Accepting payments from customers required merchant accounts, payment processors, legal agreements, and engineering work to connect it all. Publishing content and getting paid for it required a platform deal, a label, a publisher, or a network. That’s largely gone. YouTube, Substack, Patreon, and dozens of similar platforms let creators get paid directly by their audience. Stripe and its equivalents let anyone start collecting payments from customers in an afternoon — no merchant account, no enterprise contract. The App Store economy has been around long enough that publishing software and monetizing it directly is now a well-understood path. If you have an idea and want to build a way for customers to pay you for it, the infrastructure question is essentially solved.

What’s still challenging is generating demand — finding the customers, earning their trust, converting them and keeping them. That part hasn’t gotten easier. It may have gotten harder, because more people can now build and sell, which means more competition for the same attention. But the infrastructure barrier is gone.

I can see this more clearly from outside a company than I could from inside one. No legacy roadmap to protect, no internal politics to navigate, no quarterly plan constraining what you can see or build. The production barrier has collapsed. The problem of finding and keeping customers hasn’t. That gap — easier to make things, still hard to sell them — is where everything interesting is happening right now.

What’s Emerging

The question is what people are actually building with those removed barriers. What’s emerging isn’t what you’d expect.

Two models are worth naming.

Direct-to-audience businesses. Individual experts and small teams building owned relationships with the people who pay for their work — content, software, or both. No gatekeeper. No VC. No platform deal required to reach an audience.

The solo version is a Substack newsletter, a YouTube channel, an indie software tool with a Stripe checkout. The cooperative version is starting to emerge and become viable. Defector Media started as a worker-owned journalism cooperative and became a proof of concept — editorial quality, a fraction of legacy headcount, profitable. But the more significant development is what they’re doing next. Through start.coop, they’re building shared services infrastructure for small, independent news organizations: legal, HR, bookkeeping, the operational overhead that previously made running a small media collective nonviable. The back-office burden that kept these organizations from sustaining themselves can increasingly run on shared systems.

That’s a different claim than “Defector succeeded” as a unique example. It’s the beginning of a replicable model.

— co-founder and former CEO of Kickstarter — is working on the legal infrastructure side of a similar problem. The Artist Corporation, or A-Corp, is a proposed new legal entity designed specifically for creative people: shared ownership, pooled income, collective IP rights, and access to benefits, all under one structure built for how creative people actually work. It’s the collective model made legally coherent, without hiring lawyers to construct it from scratch each time. In an economy where more and more knowledge work is being treated as creative work — where the output is content, code, strategy, or expertise rather than widgets — the structure applies more broadly than its origins suggest.

New career paths. One person or a small group, configured to own their income rather than depend on it — multiple streams, no single employer. The solo version is the portfolio career — builder, advisor, fractional operator, paid writer, some combination of all of them. The collective version is a small group of senior operators with complementary expertise sharing infrastructure and working together without needing an organization built around them.

Neither is new. The experienced executive who consults, advises, and takes board seats has existed forever. Small professional partnerships have existed forever. What kept both at the margins was the coordination and operational overhead — scheduling, client communication, invoicing, legal, finance — that made running multiple engagements simultaneously impractical without either a support staff or a firm’s infrastructure behind you. That overhead is getting cheaper. AI agents can now handle much of what used to require dedicated headcount — the scheduling, the invoicing, the coordination that consumed billable time. The coordination tax that made the portfolio approach exhausting is lower than it’s ever been.

There’s also a supply-side argument. If more people are building smaller, leaner direct-to-audience businesses — which the first model suggests — there may simply be more demand for experienced operators who don’t need or want a full-time title. A world with many more small companies is also a world with more fractional roles. The portfolio career and the knowledge worker collective stop being workarounds and start being the point.

Both models are more possible now than they’ve ever been, for the same reason: the operating overhead that used to make them impractical can increasingly run with AI agents handling more of the operational load. StrongDM, a security infrastructure company, created a dedicated three-person AI team in 2025 with one rule: no human writes code, and no human reviews code. Agents write it, test it against hidden scenarios, and ship it — with the humans designing the specifications and architecture. The pattern is consistent: revenue milestones that once required hundreds of employees are now being reached with dozens. Lovable, a software creation platform, reached $100M in annual revenue with just 45 people — one of the highest revenue-per-employee ratios ever recorded for a software startup. The implication, as strategist B. Jones has argued, isn’t just that teams can be smaller — it’s that the coordination cost of being the wrong size just increased by an order of magnitude. These aren’t anomalies. This is what’s possible now — and for new companies, it will soon be standard.

But it’s not guaranteed. Most attempts will fall short. Finding and keeping customers was always the hard part, and it still is. That’s where most of these attempts will run aground. Knowing that going in isn’t pessimism. The constraints are the beginning of a real strategy.

We’re in the Clearing Again

I don’t have this figured out. I’m building through it — writing, making things, paying attention to what’s changing and what isn’t.

What I do know is this: the companies that used to offer long arcs of employment are contracting, merging, and restructuring in ways that don’t reverse. Roles are chapters now, not careers — and that’s true for most people in media and technology, not only for those of us currently in a transition. That’s not cynicism. It’s the condition anyone building a career in this market has to plan around.

The question isn’t whether to build something. It’s whether to start before you have permission or a guarantee.

Web 2.0 wasn’t announced. It emerged from the clearing — built by people who lost their footing in the dotcom bust, saw the promise and opportunity of the new technology, and started making things before anyone had a name for what they were building.

We’re in the clearing again. And what comes next — it’s always been the best part of the story.

Leave a comment

References

From this publication:

• The Labor Market They Aren’t Measuring — March 2026

• Nobody Else Knows What They’re Doing Either — March 2026

• You’re In the Creator Economy Now — April 2026

• The Writer’s Room — April 2026

Cited externally:

• Org Design in the Age of AI — Robonomics (Altimeter Capital), April 2026

• What Replaces the Org Chart — Craig Hepburn, April 2026

• Why Artist Corporations? — Yancey Strickler, 2025

• Shared Media Services Project — Defector Media / start.coop, 2026

• start.coop

• StrongDM Software Factory — StrongDM, February 2026

• Lovable reaches $100M ARR with 45 employees — TechCrunch, July 2025

Further reading:

• From Bike Shops to Bylines — Press Forward, 2025

• 45 People, $200M Revenue. The Question Nobody’s Asking About AI and Your Team Size — Nate B. Jones, March 2026

I write with AI. When writing an article, the work is what I reject.

AI produces drafts that come back coherent but voiceless. Framings that are plausible but not quite right. Arguments assembled correctly that don’t say anything worth saying. What you do with those — whether you accept them or push back, redirect, override — is where the writing actually happens.

I’ve started thinking of myself less as a writer in this process and more as an editorial director.

The ideas are mine — the questions, the hypotheses, the things I’ve noticed that feel worth saying. Those come before AI is involved at all. What I’m directing is how those ideas get shaped, challenged, and expressed. That distinction matters to me, and I think it matters for how you read what follows.

This is my twelfth article on Substack in two months, every one produced in collaboration with Claude. I’ve built a set of custom tools — Claude Skills in a private GitHub repository — specifically designed to make that process consistent and rigorous. When people assume AI-assisted writing means the AI writes and the human prompts, my experience runs the other way. I interrogate the drafts, redirect them, override them. I wrote about this in a piece on gravitational pull and escape velocity — uncritical AI use pulls work toward competent and generic. The tools I’ve built are designed to push back against that.

And I almost never accept what Claude gives me.

Effective and Efficient Are Different Problems.

The conventional use case for AI writing tools is efficiency: produce more, faster, with less effort. That wasn’t the problem I was trying to solve. I needed to produce work that was stronger — more defensible, more specific, a clearer expression of my actual thinking, more signal, less noise. Those are different problems. They require different solutions.

I’ve tried keeping a journal or a blog before. Several times, over many years. I’ve failed every time — not for lack of ideas, but for lack of a system that made consistency possible and made the output worth the effort. The question I kept running into wasn’t “do I have something to say?” It was “how do I say it in a way that holds together and that I’m proud of?”

My answer was to build tools around the process I was already doing. I was already asking AI to read a paragraph as a CPO would, or to find the arguments that wouldn’t survive a hostile read. The insight was that I could turn that into a repeatable system — run it the same way every time, stop rebuilding it from scratch in every session — and something that got better with every article I published.

That’s what I built. It’s not an “easy button.” It’s a system for doing the hard work.

The Cast of Critics

The first thing I built was what I call the Review Board — a Cast of Critics. Named personas representing the people whose judgment matters most: a CPO evaluating whether the argument demonstrates real product thinking, an executive recruiter reading for professional positioning, an editorial executive stress-testing argument rigor, a communications strategist asking what the worst-case read is.

In adversarial mode, their job isn’t to improve the piece. It’s to find what fails.

The sequencing matters as much as the personas. The adversarial pass runs before the polish pass. Always. If you polish an argument before testing whether it survives scrutiny, you’ve invested effort in something that may need to be rebuilt from scratch. Running that stress-test on an outline, before any prose exists, is where structural problems get caught early.

This piece went through that process. The adversarial panel caught a critical framing problem: I was leading with voice and copy-editing details rather than the substantive claim about argument construction. The piece you’re reading is different — and more defensible — because of that catch.

The clearest example from my published work: Nobody Else Knows What They’re Doing Either, a piece I wrote during the job search about navigating professional uncertainty. The adversarial pass caught it immediately: this could read as a laid-off executive rationalizing his situation. “Sour grapes” was the phrase. I hadn’t seen it. The piece that published had a different opening, a different proof structure, a different framing of its central claim. More defensible. Truer to what I actually wanted to say.

Professional writers have always had rooms like this. Showrunners have writers’ rooms. Public figures have communications teams. Executives have advisors who will tell them what they don’t want to hear. AI makes something equivalent available to one person working alone. My Cast of Critics isn’t fixed — it has standing members who show up for most pieces and others who get brought in or excused depending on the topic or the artifact. I’ve run the Cast of Critics on articles, LinkedIn posts, project briefs, and even other Claude Skills. Each session starts by naming who’s in the room. Each session ends when the panel stops finding things that matter.

The Voice Fingerprint

The second thing I built was an Editorial Voice skill — a style guide built from my own published work.

The hardest thing to demonstrate about this skill is exactly what makes it valuable. It doesn’t produce something I can easily measure. What it catches is the gap between “sounds like competent writing” and “sounds like me” — real, but hard to show as a before/after. It catches the places where a draft sounds fluent but generic, where the sentences hold together but my voice has left the room.

What it captures isn’t my writing itself. It’s the patterns in my writing: how I structure an argument, where I put personal moments, how I connect a broad claim to a specific detail. Building it required me to name things I’d been doing without thinking — which turned out to be useful on its own.

You can’t build a system around something you haven’t articulated.

That process changed how I write, not just how I edit.

The evidence is the response from readers. Colleagues sharing articles with people they think should read them. Hiring managers referencing specific pieces in conversations. People who’ve read earlier pieces recognizing these articles as mine. That’s what the skill protects — not elegance, but credibility. The difference between content that could have come from anyone and writing that could only have come from me.

The Publish Gate

The third thing I built was a Publish Gate — a pre-publication check against a rubric I wrote for myself. Ten criteria. Three non-negotiable gates: authenticity, earned wisdom, and what I call the AI dismissal test.

The AI dismissal test is the one that matters most for my situation. The question isn’t whether AI was involved in producing this piece. The question is: what in this piece proves my irreplaceable contribution? Not “my experience is evident throughout” — that’s not an answer. The answer has to be specific: this paragraph where I describe something I actually built, this data point from something I actually measured, this admission that cost something to put on the page.

The gate is not a rubber stamp. It can return HOLD. The first time I used it on What You Do When Things Go Wrong — And When They Go Right it came back “Ready with Caveat.” The caveat was precise: the piece had two honest personal moments, but they weren’t equal. One side of the framework had a specific owned story. The other was acknowledged but not inhabited. That pushed me to find the missing moment — a manager quote that stuck with me: “It might not be your fault, but it is your problem.” That line, and the story around it, didn’t exist in any prior draft. The gate found this gap and pushed me to go deeper. The piece that published was more complete because of it.

I’m two months in with Substack. Thirty-seven subscribers. It’s early, and I know it. I wrote about why I’m doing this in You’re In the Creator Economy Now: in a job market where the standard playbook — polish the résumé, apply to jobs, wait for callbacks — no longer works the way it once did, building a public body of work is one of the few things that still generates real signal. The subscriber count isn’t the measure. The conversations are.

The referrals. The hiring managers who’ve read a piece and reached out. The colleagues who’ve shared something I wrote with their teams and their bosses. The job interviews where I’ve followed up with a link to something that goes deeper on a topic we’d discussed. That’s what moves the needle, even when it’s hard to track.

The system didn’t make me faster. It made the work more rigorous. Every piece goes through roughly ten rounds of substantive revision — argument challenges, stress-tests, voice calibration, a final quality check. The result is stronger than anything I’d have written alone, and it’s the first time I’ve sustained a writing practice longer than a few weeks.

There’s a version of AI-assisted writing that produces a lot of content with minimal human involvement. It’s out there. Fluent, plausible, and indistinguishable from everything else. What I’m trying to build is the opposite — a system where AI amplifies the challenge and critique, and I make every editorial call. The goal isn’t to remove myself from the process. It’s to produce a stronger body of work — and I think that distinction is worth drawing clearly, whatever tools you’re using.

The gravitational pull toward the mean is real. The tools that make content easier to produce are the same tools that make it easier to produce work that isn’t actually yours. The question isn’t whether to use AI. It’s whether the output is worth putting your name on it.

Leave a comment

Two media ecosystems have been running in parallel for thirty years. One built by professionals — studios, editorial brands, institutional distributors, a value chain with a solid foundation going back a century. I walked in it every day at NBC in the halls of 30 Rock, one foot in the legacy apparatus and the other planted in the new ecosystem as I helped legacy media companies navigate the transition. The second ecosystem grew from the ground up by individuals — creators, digital platforms, algorithms that decide what audiences see.

For most of those thirty years these ecosystems occupied different lanes. That’s over. They’re now directly competing for the same finite pool of human attention, on the same platforms, by the same rules. And the rules heavily favor the new creator ecosystem.

laid out why in a keynote at the Television Academy AI Summit in March. Shapiro writes The Mediator, one of the sharper Substacks covering the business of media, and he’s finishing a book on this for MIT Press. His explanation comes down to two disruptions.

The first: digitization created a universal language for media — bits — and collapsed the cost of distribution to near zero. Moving content across the world became functionally free. That single shift produced thirty years of accelerating turbulence: the death of physical retail, the collapse of the newspaper business model, the rise of streaming, audiences fragmenting across an ever-expanding landscape of platforms.

The second disruption is arriving faster than most of us realize. Generative AI is collapsing the cost of creating content the same way the internet collapsed the cost of distributing it. The cost to make bits is heading toward zero. Which means we’re heading toward a world of effectively infinite content supply competing for a finite, essentially flat pool of human attention.

Here’s what those two ecosystems actually look like — the one I spent my career in, and the one that grew up alongside it.

The Historical Media Value Chain is the world I know well. Professional studios, editorial brands, institutional distribution. A value chain with clear roles: creatives made things, publishers packaged them, aggregators and distributors carried them to audiences. The Emerging Media Value Chain is what grew up alongside it — the creator economy. Platforms that handed anyone with a camera or a keyboard direct access to a global audience. YouTube, TikTok, Substack.

What Shapiro was describing — and what I’ve watched happen from inside the industry — is that these two ecosystems now compete for the same finite pool of attention, on the same playing field. A documentary from a major studio competes with a creator on YouTube who figured out something his audience cares about deeply. An editorial newsletter from the NYTimes competes with a solo writer who has spent three years earning her subscribers one at a time. Professional content doesn’t get preferential routing anymore.

Shapiro’s phrase for what happens next is “more cowbell” — a nod to the SNL sketch where the only solution is more of the thing that’s already too loud. His point: AI doesn’t introduce new structural problems for studios, publishers, and broadcasters. It turns up the volume on every problem that was already there. More content competing for the same attention. More fragmentation. More noise. The signal-to-noise ratio gets worse across every channel.

The cowbell is too damn loud.

This Isn’t Just a Media Problem

Media is where disruption arrives first because content is the product. When the cost of creating and distributing content collapses, you feel it immediately, structurally, everywhere. But the same collapse is now hitting the knowledge worker economy — and it’s moving fast.

The knowledge worker labor market has been quietly freezing for the better part of two years — job openings down nearly 50% from their peak, hires at 2013 levels, AI-cited cuts growing as a share of total layoffs. I wrote about the data behind this in March — before it became a more widely discussed story. What I didn’t write about then is why the standard response — polish the résumé, apply to job postings, wait for a callback to interview — no longer works the way it once did.

AI is collapsing the cost of producing professional output the same way it collapsed the cost of producing media content. Analysis. Writing. Code. Strategy documents. Research synthesis. The cost of generating all of it is approaching zero. And when production costs collapse, the core product becomes a commodity.

For knowledge workers in transition — people between roles, trying to surface new opportunities, trying to stay visible in a market that moves fast — this has produced a specific set of structural problems that map directly to what’s happening in media.

The résumé is commodity content. Creation cost: zero. Supply: effectively infinite, and increasingly AI-generated. The cover letter is commodity content. The LinkedIn summary is commodity content. And it doesn’t stop at the documents — the professional content layer built on top of them is collapsing too. Scroll LinkedIn for five minutes and count the posts that feel like they came from a real person with a real perspective. AI has made it trivially easy to generate polished, thoughtful-looking professional content at volume. When everyone can do that, none of it is a signal anymore. It’s just more noise. More cowbell.

Job postings are a broken distribution channel. This isn’t new — they’ve always been a noisy channel, and referral networks have always dominated senior hiring. What’s changed in the last two years specifically is the volume. AI application tools have flooded posted roles with hundreds or thousands of applications where there used to be dozens. Some companies have stopped posting roles publicly, or post them only after a candidate is already identified through the network. The channel that was supposed to advertise an opening and attract competition from qualified candidates has degraded further, faster.

And sitting between the candidate and the opportunity, in many cases, is an AI intermediary — an applicant tracking system running its own ranking algorithm — that the candidate doesn’t control and can’t fully see. It is same position that social media platform algorithms occupy between a creator and their audience.

The structural dynamics are the same. The arena is different.

Nobody Chose This

Nobody who is between jobs and trying to find the next one chose to become a solo creator. That’s not the plan. The plan was to do great work, build a strong network, maintain a solid presence, and trust that reputation and relationships would surface the next opportunity when the time came. Being pushed into the attention economy, and competing for visibility the way a YouTuber competes for views, isn’t something most job seekers are ready for. But that’s what this market now requires.

The same forces that commoditized media content have commoditized professional signal. Everyone has a strong LinkedIn profile now. Everyone has a polished online presence. Everyone can generate a thoughtful-looking post about industry trends in twenty minutes with AI assistance. The credential gap has collapsed. The presentation gap has collapsed. What’s left is harder and slower: an actual body of work that demonstrates real thinking, and an audience of real people who chose to follow you specifically. A track record of showing up consistently, in public, with something worth paying attention to.

This is the creator playbook. Displaced knowledge workers are being pushed toward it not by aspiration but by the same structural forces that pushed media companies to compete for attention they used to take for granted.

The Scarce Thing Is the Only Thing That Matters

Shapiro’s approach to disruption: when an input becomes abundant, the core product becomes a commodity, and value migrates to what’s still scarce. He calls these the “complements” — the things that become more valuable as content becomes cheaper — and the “chokepoints” — the structural positions that control the flow of value, and therefore allow whoever holds them to capture it.

His examples from media: when recorded music became functionally free, value shifted to live performance and merchandise. When digital photography collapsed the cost of taking photos, value shifted to editing software, cloud storage, social networks. The abundant thing gets cheaper. The scarce things alongside it get more valuable. Whoever owns the scarce thing captures the value.

For knowledge workers, the same pattern. When credentials and polished professional output become abundant and AI-generated, value shifts to what can’t be manufactured at scale — and specifically to what’s visible enough that the right person can find it.

Human provenance. Real judgment, real backstory, real process. The thing that has to come from an actual person who has actually done something. You can’t prompt your way to genuine domain experience. You can approximate it — and approximations now look pretty convincing. But there’s a difference between someone who has made hard calls in a real organization and someone who has synthesized what hard calls look like. That difference is getting harder to read from a résumé. It’s easier to read from a body of work.

Build that body of work consistently, and two things follow.

Earned direct audience. People who opted in to hear from you specifically. Not the algorithm’s audience — yours. A channel you own that sits between you and the people who might open a door for you, with no intermediary deciding whether your message gets through on any given day.

Trust at human scale. Not thousands of followers. Not virality. Enough consistent signal, over enough time, that the right people recognize your thinking when they encounter it. That recognition is what generates the warm introduction, the inbound conversation, the “I thought of you for this” message that never comes through a job posting.

I want to be clear about how this actually works, because I’ve seen this argument made too loosely. Building a direct audience doesn’t cause a job offer. The causal chain is narrower. A publishing presence keeps you visible to your existing network during a period when you’d otherwise go quiet. It generates profile views and inbound conversations that surface opportunities not posted publicly. It demonstrates the judgment a hiring manager would be trying to assess in an interview — before the interview happens. It’s a signal amplifier for the network you already have. Not a replacement pipeline.

I’m aware this is a convenient argument for someone in my position to make. I’d make it anyway, because I watched these same forces reshape the industry I spent a decade in before they reshaped the job market.

What I’m Building While I Figure This Out

I’m eight months into a job search that started when Comcast spun off its cable networks and I found myself without a company behind me for the first time in nearly a decade. I have strong hypotheses about what works in this environment. Here’s what I’m testing.

A Substack as a direct channel to the people I want to stay connected to. An editorial practice that uses AI as a collaborator — to stress-test arguments, accelerate drafting, pressure-test ideas before they’re public. Audience development tools and systems to find and engage the right readers. Each one is something I own: a direct list, a published body of work, a consistent presence that compounds over time.

This is slow. It doesn’t replace networking or direct outreach in the short term. A Substack is not going to land me a VP of Product role by itself. But it builds in a way the other channels don’t. And in a market where the fast channels are increasingly broken — or gameable by anyone with access to the same AI tools I have — the slow channel is starting to look like the durable one.

That’s not just a hypothesis anymore.

The Bet

Shapiro’s bottom line: more competition, more complexity, new opportunities — for those who position themselves correctly.

For knowledge workers navigating displacement, the translation is the same. The forces that democratized professional content creation also commoditized it. The same disruption creates both the problem and the opening. The tools are available to everyone. The discipline to use them strategically, consistently, over time — to build something real rather than just producing output — that’s what’s scarce.

The goal isn’t to become a creator. It’s to become findable, credible, and present when the right opportunity is looking.

Several pieces have led directly to referrals. People have told me they’ve shared articles with colleagues. And in a final-round interview this week, I was asked about product velocity — a topic I’d already written about. I sent the piece afterward. It did work that a résumé couldn’t do: it showed I was already living inside the problem, not just prepared to talk about it.

That’s how it works. The signal compounds before it converts, and you can’t always draw a straight line from a piece of content to an opportunity. But the line exists. I’ve seen it. The effort is building something real — and the only way to find out how far it goes is to keep going.

Leave a comment

Doug Shapiro’s keynote and slides are available on his Substack, The Mediator: dougshapiro.substack.com

The Calm app sent me the greatest push notification I’ve ever received. It stopped me cold.

“Psssst. No one else knows what they’re doing either.”

It felt like it came from an apocalyptic fortune cookie. Was it supposed to be calming? If that was the goal, it definitely failed.

This push notification became a running joke between me and my team for years. I’d send it in Slack when someone was spiraling about a decision, or when something crazy was happening. Spoiler: working in breaking news, something crazy is always happening. Sometimes I’d drop it in during one-on-ones with my team, or bring it up when the room felt stuck and nobody wanted to admit it. I’d always properly attribute the quote to the Calm app. Because it’s funny — and because it’s uncomfortably true.

What the notification is trying to do is soothe imposter syndrome. Don’t worry, you’re not alone in feeling lost. What it actually says is something much more unsettling: nobody has answers. Nobody has a plan. Everyone — including the people you assume have it figured out — is really just winging it. That’s not comforting. It’s distributed anxiety.

The Pedestal Doesn’t Exist

But sit with it for a second.

If nobody knows what they’re doing, then why should I assume anyone else has things figured out? We scroll LinkedIn or Instagram, and see people who look like they have all the answers. Confident posts, polished takes, impressive titles. And we think to ourselves: they know something we don’t.

Do they? Or are they just performing?

The notification cuts through imposter syndrome — not by reassuring you that you’re good enough, but by revealing that the pedestal you put other people on doesn’t exist. The gap between you and the people you admire is smaller than you think. Often it’s not a gap at all.

If nobody knows, then the question isn’t am I qualified enough?

Instead, it’s: why not me?

The Map No Longer Matches the Territory

We’re living through one of the great reshuffles. AI is upending long-held assumptions faster than most people, including most leaders, can process. 5-year plan? Forget it. Annual planning? Nope. Everything is being reconfigured and recalibrated every couple months, it seems. That’s how fast AI is moving. And it’s not slowing down.

I spent nearly a decade building digital products at NBC News, MSNBC, and CNBC. Something I thought I understood deeply is being rewritten in real time.

How we build products is being fundamentally reconfigured. For most of my career, the canonical product team looked like this: a product manager, a project manager, a tech lead, four or five engineers, a designer, and a QA lead. Eight to ten people, synchronized on two-week sprints, quarterly OKRs. That structure wasn’t arbitrary. It was sized for the coordination costs of human collaboration. I ran multi-team portfolios built on this model for years. It was how things worked.

It’s not how things work anymore. A PM who can build a prototype in the morning, run data analysis by afternoon, and ship a research synthesis without a dedicated analyst doesn’t need an eight-person team to validate ideas. I’ve been building AI tools solo for the past year. I’m not estimating story points or managing sprint capacity. I’m making decisions about what’s worth building at all, and then moving. The constraint isn’t effort. It’s strategic clarity.

A lot of experienced people built their careers on pattern matching, and those patterns are crumbling. When everyone has access to the same AI tools, the outputs converge. Velocity increases. But product sense — knowing what’s worth building and for whom — still matters a lot.

People who’ve accumulated genuine wisdom still matter — maybe more than ever. But expertise alone isn’t enough when the territory no longer matches the map. The most interesting space now is where real experience meets willingness to unlearn and throw out the old playbooks given the new game.

The Badge Is Not the Skill

Losing a title or a role can feel like losing the thing that made you credible. You had a company behind you, a team around you, a position that told people who you were. Without it, imposter syndrome doesn’t just creep in — it moves in and rearranges the furniture.

I’m writing this as someone in transition. You can take that however you want. But I’d argue that being outside the corporate machine right now, uncomfortable as it is, means I’m building with the new tools firsthand instead of reading about them in someone else’s strategy deck.

The antidote to imposter syndrome isn’t reassurance. It’s action. Not frantic activity, but moving with strategic intent and court awareness. In basketball terms — keep your dribble alive, head up, and see the whole floor.

Start making or building the thing you envision. That doesn’t mean start a company (although for some people that may be the right move). What it might mean is to write something that reflects your actual thinking. Build something that demonstrates the skills you’d bring to your next role. Show up as the leader you want to be — before someone gives you permission to be one.

That’s what I’m trying to do. I’m building AI projects each month — the full cycle from discovery to build to beta testing with real users, iterating based on feedback. For one example, I built a civic data tool for NYC parents — no team, no budget, no title — using AI to do work that would have required a full engineering team a year ago. Journalists engaged with it. Parents used it. The skills don’t disappear when the badge does.

The distribution playbook is being rewritten too. Social platforms used to be a top referrer to our sites at NBC. That traffic disappeared a few years ago. The platforms walled off the audience for themselves. Content still lives on social media, but the algorithm stands between you and the audience you’ve built on the platform. I see this on LinkedIn every day — I can’t reliably reach my own network. Search, another important traffic driver to publisher sites, has dropped by more than a third since Google added AI summaries. AI tools like ChatGPT aren’t referring traffic either, although many hope this will change. The old audience development playbook is dead. What’s replacing it is earned audience: loyal subscribers who chose you directly. That’s a much harder and slower journey. But it’s more durable once you build it.

That’s the bet I’m making with this Substack, starting from zero with no institutional brand behind me. It’s the hardest problem I’ve worked on. And it’s the next one I’m writing about — because most of the old playbook is already gone.

Nobody knows what they’re doing. That’s true. It’s terrifying and it’s liberating — sometimes in the same breath.

Here’s my approach: Keep building. Keep showing up. Stay in motion with intent. Head up, eyes on the court.

If there’s something you want to build, write, try, test, make — why not you? And why not now?

What’s actually in your way?

Leave a comment

There’s a question that comes up frequently in senior leadership interviews:

What kind of leader are you? What kind of culture do you build?

I’ve answered this question a lot lately. I’m in the middle of a job search, which means I’ve had more of these conversations in the past several months than in the previous several years combined. And the more I’ve given it, the more it’s come into focus.

I talk about being a player/coach — someone who is in the work with their team, not above it. And I talk about two things I try to hold at the same time: playing to win and playing the game well.

Playing to win means knowing the business impact you want to create and going after it. That part is table stakes. Every leader says some version of it.

Playing the game well is something different. Process and tactics matter — how you break down work, how you prioritize, how you communicate with senior executives and stakeholders. But those things sit on top of something more foundational. Playing the game well is about the attitude and behavior you bring to the work — and whether you can inspire those behaviors and make them contagious across your team and throughout your organization.

I wrote about this foundation before — in a piece about The Mindset Underneath the Work, built around four videos I shared with my team during our annual reviews: Giannis on failure, Kara Lawson on hard things, Fred Rogers reframed, Bruce Lee’s “be water.” That piece was about internal orientation — how you show up inside yourself. This one is about how that orientation shows up in your behavior toward others.

You Already Know Which Side You Want to Be On

I’ve been working through this as a simple diagram. Two grids — call them the Wrong Frame and the Right Frame. Four situations mapped across both: bad things happening to you, good things happening to you, bad things happening to others, good things happening to others.

Here’s what the Wrong Frame looks like:

• Bad things happen to you → point fingers

• Good things happen to you → take credit

• Bad things happen to others → blame them

• Good things happen to others → jealous sandbagging

And the Right Frame:

• Bad things happen to you → take accountability

• Good things happen to you → share credit

• Bad things happen to others → help

• Good things happen to others → celebrate

Look at that and you know exactly which side you want to be on. Every leader does.

That’s the problem. This looks obvious. But in practice it is not easy.

The Wrong Frame isn’t where bad leaders live. It’s the default — the path of least resistance when things get hard, in competitive environments, in organizations where the incentives are running the wrong direction. The Right Frame is what you have to actively choose, repeatedly, under conditions that consistently make the other choice easier.

Knowing What to Do and Actually Doing It Are Different Things

When something goes wrong and fingers are pointing, the instinct to defend is immediate. I’ve been in situations where I’ve felt it — the frustration, the unfairness, the urge to point at everything that contributed to the failure that wasn’t mine. That’s human. What I’ve learned is that the feeling needs somewhere to go — a private conversation, a trusted advisor, a moment to process before a better response takes form. The reframe isn’t about suppressing the reaction. It’s about making sure that reaction doesn’t become your public behavior.

A manager said something to me once that I didn’t love hearing in the moment: “It might not be your fault, but it is your problem.” My first read was defensive — why is this my problem alone? But that wasn’t what he meant. He meant that everything within my purview — my teams, my projects, the external forces landing on us — was mine to own. Not to blame myself for. To show up for. To be accountable for solutions, not just causes.

That connected to something I wrote about in The Mindset Underneath the Work: it’s not your problem or their problem, it’s just the problem. Move from MY problem to THE problem — and something unlocks. Blame has nowhere to land. The defensive energy dissipates. People can actually start working on it together. There’s a time for root cause analysis, but when you’re fighting a fire you first need to put out the fire. Blame while the fire is burning makes the fire worse and the blame land harder.

The same dynamic plays out on the other side of the framework when good things happen — and it’s just as damaging. Leaders don’t fail to celebrate because they’re malicious. They fail because they’re busy, stressed, and focused on what’s next. The absence accumulates invisibly — and by the time morale shows up damaged, it’s been building for months.

I’ve been called out for this before. I used to give broad team-wide thanks after a big push — “great work everyone, really proud of what we accomplished” — and I meant every word of it. What I didn’t understand then is that non-specific recognition can land as disingenuous, even when it’s genuine. The people doing the work know that senior leaders often don’t have visibility into who did what. A blanket thank-you can feel like confirmation of that distance rather than a bridge across it.

What actually motivates people is specific, accurate praise. A VP who knows that a particular engineer solved a problem nobody else could crack, and says so by name, lands completely differently than “the team did great work.” You can’t always thank everyone, and some people don’t want to be singled out publicly. But a private note with a specific observation costs almost nothing and matters enormously.

The Tone You Set Becomes the Environment Your Team Works In

These behaviors aren’t just a reflection of individual character. They’re a reflection of conditions. A team that feels blamed will become defensive. A team whose credit gets absorbed upward will feel unseen and lose motivation. A team that isn’t celebrated will stop celebrating each other.

As a leader you are shaping those conditions whether you intend to or not. The behavior you model sets the conditions for what others feel permitted to do. A leader who takes accountability publicly — who says “that was my call and it didn’t work, here’s what I’d do differently” — gives their team permission to do the same. People stop managing their exposure and start solving the problem. The inverse is equally true and faster: one visible deflection of blame can reset months of trust in a single meeting.

Most of the time you won’t be making a conscious choice between the two frames — you’ll just be reacting. The work is catching yourself. Noticing when the defensive response is right there, easier and more satisfying than the alternative, and asking: what would this look like if I chose the other path?

That question, asked often enough, changes what your team feels permitted to do.

It looks obvious until you’re in it.

Leave a comment

All of it, every sprint, on the same backlog.

It didn’t fail dramatically. Nothing caught fire. What happened was difficult to diagnose and probably feels familiar to those of you leading a product portfolio at scale. Simply put, we couldn’t do everything. So we made tradeoffs. And the tradeoffs followed a predictable pattern. Business objectives tended to win. Engineering and operational goals (the work that didn’t have a revenue metric attached) tended to lose. Not because leadership was wrong to prioritize growth. Because nobody had a language to make a convincing case for the other side.

Then a compliance mandate would land. The kind you can’t defer. And suddenly the work that had been losing every prioritization fight became “must do,” and it crowded out everything else, including the business goals that had been winning, and were still expected by our leadership. We got into a cycle where we ended some quarters attempting to do it all but doing none of it well. Teams were stretched across too many priorities. The surface area felt too large and we exceeded our capacity on everything, delivering on nothing cleanly.

There was no quick fix for team capacity. We simply couldn’t add head count, nor would it have helped in the short term. But there was a structural problem, and I needed a vocabulary to describe it clearly enough to argue for a better solution. This problem isn’t unique to media — any product org managing a complex portfolio across multiple teams and lines of business will recognize it.

That’s where the Whole Product Flywheel came from.

Engineering priorities need to play on the same field as business priorities

“Technical debt” might be the most expensive phrase in product development. Not because the problem it describes isn’t real (it is), but because the language guarantees the wrong outcome. You walk into a planning meeting, you say “we need to pay down technical debt,” and you’ve already lost. The word “debt” implies someone made a mistake. Nobody budgets to fix mistakes. They budget to grow the business.

What if you called it what it actually is — a deferred investment in the infrastructure that enables everything else you’re trying to do?

Same work. Completely different organizational psychology, and a completely different conversation in a planning room. Engineering teams that can make a business case for “Enabler” work win more often in prioritization fights. Teams that can only confess to accumulated debt won’t get prioritized.

The Flywheel is the structure that makes the business case possible.

Product leaders are responsible for more than your typical roadmap measures

I’m not going to claim this is a new framework. It isn’t. The components — DORA metrics for engineering health, OKRs for goal alignment, platform team models for ownership design — have been in circulation for years. What the Flywheel adds is a unified vocabulary that makes all of those components legible to the same room at the same time. That’s a different thing from a prioritization framework. It’s a translation layer. And it turns out that translating priorities into a common language is the key problem most product orgs are actually dealing with.

The Whole Product is made up of two sides, Product Outcomes and Product Enablers, and four quadrants that describe the full priority surface area of what a product organization is responsible for.

Product Outcomes are the things your business and your users can see directly:

• User Experience / Customer Experience (UX/CX): Features, usage, engagement, customer success, satisfaction. The product your users interact with.

• Growth: Distribution, acquisition, activation, retention, monetization. The engine that expands and sustains your audience or customer base.

Product Enablers are the things that determine your capability to deliver those outcomes:

• Developer Experience (DX): Team collaboration, momentum, quality, automation, supportability. How effectively your engineering teams can build and maintain the product.

• Operational Excellence (OX): Releases, stability, performance, resiliency, security, privacy, analytics, partner and stakeholder support. The infrastructure that keeps the product running reliably and your org operating efficiently.

The conventional mistake is treating Outcomes as “the product” and Enablers as overhead. This is how you end up with a prioritization process that systematically defunds half of the priority surface area. Enablers don’t disappear when you stop investing in them. The decay shows up later: in velocity, in stability, in the compliance mandate that lands without warning and crowds out six weeks of roadmap.

Enabler work earns a more compelling business case when it speaks the same language as other priorities

The reason Enabler work loses the prioritization fight isn’t just language. It’s that the metrics don’t exist yet in most organizations. Feature work has conversion rates and active user counts. Growth has acquisition cost and retention curves. Enabler work has... velocity? Test coverage? These aren’t illegible. They’re untranslated. And in a prioritization fight, untranslated loses.

The Whole Product Flywheel uses a Goals → Signals → Metrics structure for each quadrant. The goal is the outcome you’re investing in. The signal is the observable evidence that you’re moving toward or away from it. The metric is the specific, measurable target that makes the case.

Here’s what that looks like for Developer Experience, the quadrant where the translation is hardest and the stakes are highest:

These metrics are illustrative — calibrate them against your actual baseline and negotiate them with your engineering leadership. One important framing note: these are team health diagnostics, not performance management tools. The moment a manager uses “PR review to approval time” to pressure engineers into rubber-stamping reviews, you’ve weaponized a health signal into a compliance metric. These numbers exist to start honest conversations about what a healthy developer experience should look like and how to get there, not to create new ways to punish teams for being thorough.

A platform team that walks into planning with a measurable hypothesis about ROI is playing a different game than one asking leadership to trust that their work matters.

A shared ROI vocabulary made a smart reorganization possible

The framework gave us a shared vocabulary. What it made possible was a structural argument.

I used it to make the case for splitting and re-orienting our teams.

The iOS, Android, and CTV teams were each trying to hold all four quadrants of the flywheel simultaneously. That’s too much priority surface area for a single team to serve well in a single planning cycle, especially when business objectives, engineering objectives, compliance mandates, and growth priorities are all competing on the same backlog. Something always loses. And the priorities that lose consistently don’t fail visibly. They just accumulate drag that eventually becomes a constraint.

The split we made was principled, not political. We created a dedicated Enabler-focused platform team and redirected ownership of DX and OX work to them. This freed up the other teams to focus more cleanly on Features, UX and CX. And we brought in a team from another part of the organization focused specifically on subscription revenue Growth, because we were serving multiple brands and lines of business and the subscription line of business needed its own specific focus.

The outcomes weren’t perfect and they weren’t immediate. Building shared vocabulary across five distinct scrum teams takes time. Getting engineering-led OKRs into the planning conversation, and keeping them there when Growth pressure spiked, required ongoing advocacy. But over time the portfolio operated differently. We hit our OKRs consistently. We met compliance goals without blowing up the roadmap. We shipped features that grew audience engagement and launched subscription services with real business impact. We did it with a planning process transparent enough for leadership to actually see the tradeoffs being made, instead of having them happen invisibly in a backlog nobody outside the team fully understood.

Splitting teams is not a metrics decision in practice. It’s a budget fight, a headcount negotiation, and a leadership trust exercise. This framework gives you the language to have that conversation — not to make it easy, but to move it from “trust me, this matters” to “here’s the evidence, here’s the structure, here’s what changes.” That’s a different conversation and a more convincing argument.

The signal for when to split: when the same categories of work keep rolling over sprint after sprint, not because they’re deprioritized explicitly, but because there’s no protected capacity for them. Build your Enabler Goals/Signals/Metrics first, even rough, so you have a baseline to recognize that pattern when it appears. The GSM structure above is a starting point. Calibrate it to your org.

AI is accelerating the Flywheel faster than most orgs can reinvest the gains

Some engineering teams are shipping faster than they ever have. Others are drowning in pull requests generated by tools that write code faster than humans can review it. Both are using the same AI tools. That gap — between teams realizing genuine DX gains and teams accumulating new bottlenecks — is what most leadership conversations about AI are skipping over.

Done well, AI tools represent one of the most significant potential Enabler investments your org has ever encountered: faster builds, better test coverage, reduced maintenance burden, automated operations. The potential to free up meaningful capacity across the entire flywheel and redirect it where it’s needed most.

But the pressure to extract those gains for Growth is arriving before the gains have fully materialized. Leadership sees AI everywhere and draws a familiar conclusion — cut costs, push harder on Growth, move faster with the same or smaller teams. Meanwhile Growth is facing structural headwinds AI can’t fix: distribution is broken, advertising markets are contracting, AI is eating the search and social referral traffic that many Growth models were built on.

So here’s the question the Flywheel helps you answer: when your organization starts realizing AI acceleration gains, where does that capacity go? Back into the Enabler side to compound? Into UX/CX to improve the product your users actually experience? Or straight into Growth pressure, consumed before it can accumulate?

This isn’t a prediction. It’s a hypothesis worth taking seriously before the decision gets made by default, which is how most organizations will make it.

When tradeoffs speak the same language, planning becomes transparent

The framework doesn’t make hard decisions easy. Planning is still political. Headcount and budgets are still finite. Leadership will still ask for Growth first.

What it does is change the quality of the conversation.

When every part of your product has measurable goals and visible ROI — when DX has metrics that translate into business outcomes, when OX has targets that make the cost of neglect clear, when the Enabler side of the flywheel is as legible as the Outcome side — investment decisions stop being arguments about trust and start being arguments about evidence.

I’ve seen what happens when engineering teams can’t translate their goals into ROI outcomes well enough. The work without a revenue number attached loses every planning cycle, the decay accumulates quietly, and eventually surfaces as a crisis: a compliance mandate, a stability incident, a team that can’t move fast enough to capitalize on a growth opportunity because the platform underneath them is dragging.

I built this framework because I needed a language to prevent that. And because the teams I worked with deserved to have their work, all of it, understood and valued for what it actually is.

An investment. Not a debt.

The vocabulary exists. The team structure is proven. The only question is whether you treat the AI moment as a budget extraction opportunity or a flywheel investment. That decision — made mostly in annual or quarterly planning rooms, mostly without being named explicitly — is yours to make deliberately, or let happen by default.

Leave a comment

The Whole Product Flywheel framework was developed during my time leading Apps platform teams at NBC News Group, covering NBC News, MSNBC, and CNBC digital products. It draws on prior frameworks including DORA metrics and OKR structures, synthesized into a unified vocabulary for cross-team prioritization and planning. NBC News Group was restructured in 2025 when Comcast spun off its cable networks into a new company, Versant.

I keep seeing the same argument. In LinkedIn posts, in essays, in talks. It goes like this: AI can do many things, but it doesn’t have taste. It doesn’t have judgment. That’s what humans bring. That’s the differentiator.

The more I see it repeated, the more I think we’re flattering ourselves.

AI doesn’t have taste. It has gravity.

Here’s what I think is actually happening. AI reflects median human aesthetic preference at scale. It’s trained on the aggregate of everything that came before — the good, the bad, the average, the forgettable. Given context and direction, it produces work that most people find acceptable. Competent. Plausible. Clear enough. And depending on what the work is, maybe good enough.

That’s not the same as having taste. It’s something more specific — and more worth understanding.

When you use AI without interrogating it — when you accept the first plausible answer, when you let it confirm what you already think, or just get to decent acceptable outputs quickly — your work doesn’t just risk being average. It gets pulled toward the center of everything that already exists. Not toward bad. Toward the mean. Toward the middle of the entire distribution of human output the model was trained on.

This is why I see so much repetition. So much competent sameness. The apps that look identical because they’re all built with the same tools, accepting the same defaults. The essays making arguments you’ve read before, framed almost the same way. The “taste and judgment” argument itself — by now almost a meme, reproduced so many times it has stopped requiring thought.

Two things you can’t fake or shortcut.

Escape velocity is the right frame. Not a one-time push — sustained force, applied consistently, in a specific direction.

The force is interrogation. Pushback. Challenge. Redirection. The refusal to accept plausible as sufficient. The insistence on asking not just “is this good?” but “is this true?” and “is this mine?” and “why this and not something better?”

Two things produce this force and direction. Neither can be faked or shortcut.

The first is earned wisdom. Not expertise in AI tools. Expertise in something real, accumulated through actual experience and actual failure over actual time. The kind of knowledge that lets you hear a confident answer and know immediately whether it’s right or whether it is BS — not because you looked it up, but because you’ve lived in the problem long enough to feel when something is off.

Earned wisdom is what lets you say: that’s not right.

The second is insatiable curiosity. Not openness. Not keen interest. An active, relentless drive to push past the answer that would satisfy most people toward something truer, better, or unexpected. The question that keeps going even after you have something that plausibly works.

Curiosity is different from wisdom in one important way: wisdom accumulates over years. Curiosity is a choice you make every day — and you can lose it. The gravitational pull is comfortable. A plausible answer that looks finished is comfortable. Stopping there is always available. Curiosity is what makes you keep going.

Curiosity is what compels you to keep asking: but is that actually right?

Applied consistently, earned wisdom and insatiable curiosity are what generate escape velocity. What push the work beyond the mean toward something that couldn’t have come from anyone else.

What I’m actually reaching for.

I’ve been building AI-assisted projects for the past year — and writing this piece with AI assistance, which feels relevant to say here. I use it as a sounding board, a creative partner, a source of pushback. I look for the places where it challenges my thinking, not just confirms it. I provide the spark, the perspective, the context, the direction. Whether I’m doing that well enough is a question I keep asking myself.

What I’ve learned is that the results are almost entirely determined by what I bring to it. Not technical fluency. The quality of the thinking, the clarity of the problem, the specific strategic intent — and whether I keep interrogating rather than accepting.

When I accept a plausible answer, the work drifts toward the mean.

The actual work is pushing back, redirecting, asking why this and not something better. Not every push lands. But consistently enough that I’ve learned to feel the difference between producing output that’s good enough and shaping it toward something better.

I’ve tested this across a range of projects: a karaoke language learning app, an AI explorer for NYC school quality metrics, a labor market analysis, an exploration of remixing NBA highlights. What I reach for in every one is work that has a reason to exist. A specific problem. A specific audience. A point of view that can be defended from multiple angles: through the product strategy, the data, the UX, the technical choices, the ethical ones. Depth that goes beyond the surface.

I’m not claiming any of it approaches the ceiling. But each of those projects carries earned wisdom from decades building products at the intersection of media, journalism, technology, and audience — and the insatiable curiosity to keep asking whether it’s actually right, whether it actually solves something. The ceiling is still a long way up.

That’s what I’m actually reaching for: work that says something only I could say.

The right friction propels. The wrong friction anchors.

Most of what people are sharing publicly right now are the results of one person working with AI. No colleagues pushing back. No skeptical peers. No editorial layer. And too often it is merely evidence of the gravitational pull. This is why all the prototypes start to look and feel the same. Why the essays and LinkedIn posts all say the same thing you’ve read dozens of times.

But escape velocity doesn’t have to be a solo achievement. When a group of people with different expertise, different perspectives, and genuine curiosity engage with a work in progress, the accumulated wisdom and friction can propel it toward something none of them could have reached alone. The journalist who catches a false premise. The engineer who sees the structural flaw. The skeptic who asks the question everyone else stopped asking.

The same forces can also work in reverse. Groups optimize for consensus. Organizations have gravity too — hierarchy, risk aversion, the pressure to ship something acceptable rather than push toward something better. The same collective energy that can accelerate escape velocity can just as easily anchor a work to the mean.

Real authorship requires something beyond bringing the spark and holding the direction. It requires continuous discernment: knowing which feedback is propulsive and which is gravitational. Knowing when to follow a challenge that reorients the work toward something truer, and when to discard input that pulls it back toward the comfortable and familiar.

That discernment is itself an exercise of earned wisdom and curiosity. The same force, applied to the inputs rather than just the outputs.

In an organization you may not always have the agency to make these decisions. Consensus has weight. Hierarchy has weight. But even partial agency matters. Knowing which inputs you can push back on, which feedback you can redirect, which defaults you can refuse — that is itself an exercise of earned wisdom. As a solo builder, making something from scratch, the agency is entirely yours.

So is the responsibility.

The ceiling is higher than anyone is reaching.

I haven’t gotten close to it. What I can see clearly from where I am is how rarely anyone tries — how easy it is to call competent work good work, how seductive it is to stop once you have something that looks finished, how much gravity there is pulling toward the mean in every direction.

The “taste and judgment” argument is, I think, a symptom of this — and I include my own past versions of it. It’s a flattering story we reach for: humans have this special quality, AI doesn’t, therefore we matter. Repeated enough times, it stopped requiring thought. It became exactly the kind of work it was trying to argue against. We didn’t notice, because the pull was comfortable and the story felt true.

What actually matters isn’t taste as a passive quality you have or don’t. It’s whether you’re generating enough force to escape the pull — through earned wisdom, insatiable curiosity, and the friction of people who will genuinely push back.

The people who can help you escape the pull are the ones who will add force rather than weight — whose wisdom and curiosity create friction that propels rather than anchors. That friction is uncomfortable. It is also the only way to find out how high the ceiling actually is.

The next time you read an argument about how humans have taste and AI doesn’t — ask whether the person making this argument is demonstrating the thing they’re describing or just regurgitating it. Ask whether that argument achieved escape velocity or just cleared the floor.

Then ask the same question about your own work.

The pull is real. The ceiling is higher than you think. The force to push beyond has to come from somewhere.

Leave a comment

Not the goals. Not the projects on the roadmap. Those were already set. I mean the mindset for how I wanted my team to approach the work, and whether someone could actually do what we were asking of them when conditions got hard. Which, in media and technology, they always did.

Last year, I sent my team these four video clips before our individual conversations. No framing. No instructions. Just: watch these, and let’s talk. Each one was less than three minutes long. None of the clips had anything to do with product management, or media, or technology. They were basketball players and coaches, martial artists, and a man who could be your grandpa wearing a bow tie.

The moment I’m describing — a field under pressure, teams unsure of their footing, professionals trying to figure out what they’re worth and where they fit — was true when I shared these clips. It’s also true right now, for a lot of people.

I should say: I’m in the job market now after a layoff last fall. I’m not writing this from the other side of a transition I’ve already navigated. I’m writing it from the middle of the same uncertainty a lot of people are sitting in — employed or not. Which is maybe why these clips feel more useful now than they did when I was the one sharing them.

The first video: Giannis Antetokounmpo on “failure”

After his team, the Milwaukee Bucks, were eliminated from the playoffs, a reporter asked Giannis Antetokounmpo — two-time NBA MVP — whether the season had been a failure. He paused, then pushed back on the premise. He talked about steps — each one building toward something. He named players who’d failed for years before winning championships. He wasn’t dismissing the loss. He was refusing to let it be the final word.

What followed was a panel on Inside the NBA — Barkley, Shaq, Kenny Smith, Ernie Johnson — each sitting with what Giannis had just said and working through it differently. Four people who’d competed at that level, disagreeing about what failure means and what it doesn’t. That conversation was part of what I wanted my team to see. Not a verdict. A debate worth having.

The reason I sent it was simple. Product teams fail constantly. Features get cut. Launches disappoint. Strategies get reversed. Meetings with stakeholders go sideways. Annual reviews, almost by definition, involve a conversation about things that didn’t go the way we planned. I wanted my team to walk into those conversations — and into the year ahead — with a framework that didn’t make failure mean over. Failure is a step. The question is whether you take the next one.

That’s harder to hold onto right now than it sounds. Layoffs have a way of making failure feel terminal. A long job search has a way of making every rejection feel like a verdict. It isn’t. Resilience is a choice — and the next step you take is the one that matters.

The second video: Kara Lawson on “handling hard things”

Kara Lawson coaches the Duke Women’s Basketball team. The clip is short — less than two minutes — and it contains one of the more useful reframes I’ve encountered. She’s talking to her players about hardship, and she says something that sounds almost counterintuitive: things don’t get easier. They just get harder. And the only way to navigate that is to build the capacity to handle hard things.

This is the easiest it will ever be.

That sentence has a way of landing differently depending on where you are in your career. Early on, it sounds motivating. In the middle of something genuinely difficult, it sounds almost cruel. But it’s accurate — and accuracy matters more than comfort when you’re trying to prepare a team for what’s actually coming.

What I wanted my team to understand is that the answer to a harder environment is not to wait or wish for it to get easier. The only viable path is forward — building the capacity to handle what’s coming. That’s a skill. You have to develop it. It’s earned, not given. It comes from exposure, from doing hard things and surviving, from building the kind of track record with yourself that lets you believe you’ll make it through the next one.

Resilience keeps you in the game. Getting better at handling hard things moves you forward. But both of those are still about the individual — what you carry, what you build in yourself. The third clip was about something harder, and more useful: what you do for the people around you when everyone is under pressure at the same time.

The third video: Fred Rogers on “helpers”

The clip is from an interview where Fred Rogers describes what his mother told him as a child: when something scary happens, look for the helpers. It’s advice for children — a way to find safety and reassurance in moments of fear or chaos. Most people who’ve encountered it carry it that way.

I brought it to my team differently. Don’t look for the helpers. Be the helpers.

That inversion matters more than it might seem. In any organization under pressure — and media organizations have been under real, sustained pressure for a long time — there’s a gravitational pull toward defensiveness. Toward protecting your team’s work, your roadmap, your resources. Toward framing a problem as someone else’s failure before it has a chance to become yours. I’ve watched it happen to genuinely good people. I’ve probably done it myself.

The people who move differently are easy to spot, and they’re the ones who get remembered. When something goes wrong — a launch fails, a strategy reverses, a team is suddenly understaffed and under-resourced — they don’t spend energy establishing distance from the problem. They show up to it. They treat it as the problem, not as your problem or their problem. They ask what needs to happen next and they find a way to be useful.

This isn’t about selflessness. It’s about effectiveness. The people who become genuinely indispensable in difficult organizations are the ones who can set aside ego and defensiveness at exactly the moment when everyone else is putting up their guard. That’s a choice — and it’s one of the clearest signals I know of someone operating at a senior level. Not because of what they’re in charge of. Because of how they show up when things are hard.

And when someone makes that choice, something shifts in the room. I’ve watched it happen. One person drops the defensiveness, stops angling for distance from the problem, and just starts seeking solutions without pointing fingers — and suddenly the dynamic changes. The conversation that was going nowhere starts moving. Other people follow. It turns out a lot of the friction wasn’t about the problem itself. It was about everyone waiting to see who would go first.

The fourth video: Bruce Lee on “the shape of water”

The clip is one of the most quoted things Bruce Lee ever said: be water, my friend. Water takes the shape of whatever contains it. It flows around obstacles. It doesn’t resist — it adapts.

I sent this one last, and deliberately so. The first three clips are about internal posture: how you hold failure, how you build capacity, how you show up for others. This one is about how you move through a world that keeps changing the shape of the container — which is a different problem.

Product management lives with uncertainty. Unclear strategy, shifting priorities, misaligned stakeholders, roadmaps that don’t survive contact with reality. The instinct — especially for people who are good at their jobs — is to impose structure. To seek clarity. To resolve ambiguity before proceeding. Sometimes that’s right. But often clarity and certainty isn’t readily available, and insisting on it might be the thing that keeps you stuck.

Being like water doesn’t mean being passive. Water moves. Water fills the space it’s given. Water, given enough time and pressure, carves through stone. The posture I was describing to my team wasn’t shapelessness — it was the refusal to be rigid when rigidity was the actual obstacle. Adaptability as a professional practice, not as a fallback when the plan fails.

That feels especially relevant right now. The product teams I’ve watched struggle most through this period of AI disruption aren’t the ones dealing with hard problems — they’re the ones insisting the problems stay the shape they already know how to solve. Rigidity dressed up as standards or best practices. Resistance dressed up as rigor. The teams that move well are the ones that stay curious about the new container, even when it’s uncomfortable, even when it doesn’t match their long-held assumptions.

I spent nearly a decade at NBC News Group leading digital product teams — working with editorial, data journalists, engineers, designers, and executive leaders building products for some of the largest news audiences in the country through some of the most consequential news cycles in recent history. I’ve had versions of this conversation across budget cuts, failed launches, platform pivots, and the slow grind of a business model that was never quite stable enough to feel safe.

What I kept coming back to, in those conversations and in the clips I sent last year, is that the conditions we’re all operating in now aren’t temporary. They’re not a hard season that resolves into an easier one. The field is changing, the market is changing, the tools are changing — and the people who navigate that well aren’t the ones who found a way to wait it out. They’re the ones who got better at hard things, showed up as helpers when it was difficult to do so, and stayed fluid enough to be useful in whatever shape the work takes next.

I’m doing that work myself right now. Some days it’s clearer than others.

If you’ve recently had your own annual review, or recently found yourself not needing one, I hope something here lands.

Leave a comment

It also created an obsession with the NBA footage I grew up watching live. But every time I return to that footage, it's the same broadcast frame: two camera angles, somewhat grainy, not HD, definitely not 4K. Modern highlights don’t have this problem — when Jalen Brunson hit that step-back three to close out the Pistons in last year’s playoff series, the NBA posted eight camera angles before the night was over.

That asymmetry kept bothering me. As a product leader who’s spent two decades building for media companies, I recognized it as a business problem: archives full of irreplaceable moments, excluded from the modern attention economy because the footage can’t survive the speed of current feeds. A grainy clip from 1994 might be nostalgic, but it doesn’t stop the scroll for today’s fans who expect more from highlights.

I wasn’t the only one thinking about this. At All-Star Weekend 2026, the NBA demoed real-time player POV rendered through 3D avatars. ESPN went a similar direction for Christmas Day with Sony’s Beyond Sports — live player tracking converted to real-time animation. Both are impressive, but the result looks like watching NBA 2K: synthetic scenes built from data, not real footage. Fans of the video game format will find it familiar. Fans who want to feel closer to the actual game won’t.

My goal was different: unlock new ways of experiencing footage that already exists. The AI here works within the original scene, not replacing it. And unlike these other approaches, I wanted something that works retroactively on archive footage.

The first thing I did was fire up a state of the art AI model and try creating immersive views and alternate angles from broadcast frames. It went badly. That failure ended up being the most important finding of the whole project.

What I Tested

I explored three distinct approaches.

The first: perspective synthesis — generating camera angles that were never filmed. What does Brunson’s game-winning three look like from the defender’s point of view? From under the basket? These angles exist in broadcast replays for maybe 2% of modern highlights and essentially none of the footage shot before 1990.

The second: character substitution — replacing real players with AI-generated avatars. Same play, same arena, same moment. Different characters. Think Space Jam as a production mode rather than a one-off movie.

The third: 3D scene reconstruction — using Apple’s SHARP model to reconstruct actual 3D spatial geometry from a single photograph, then orbit the camera through that immersive scene. No AI hallucination. Just the spatial structure implied by the original image, transformed in 3D space.

The best perspective synthesis result — a 360° orbit around Brunson’s series-clinching shot — is genuinely compelling as a social media asset, even if you can tell it’s AI. The avatar swap works surprisingly well until the moment Patrick Ewing and Alonzo Mourning merge into a single impossible creature. And the 3D reconstruction result that stopped me was Walt Frazier’s jumper in Game 7 of the 1970 NBA Finals — an image sourced from a newspaper photo and upscaled through open-source tools, producing an immersive 3D scene from 55-year-old film photography.

Here’s what I learned.

The First Wall: Video Frames Are Useless

My first instinct was to work from actual broadcast footage. I had 8 synchronized camera angles of the Brunson shot — overhead, baseline, close-up on the release, the standard wide — eight perspectives of the exact same moment.

The results were unfortunately incoherent and nonsensical. Broadcast video frames are too degraded for these models: motion blur, compression artifacts, and telephoto lens distortion that flattens depth. My sharpest capture made Brunson appear to be standing directly under the basket — the AI model had no spatial information telling it that Brunson was actually at the three-point line, so everything it generated placed him in the wrong position relative to the hoop. Preprocessing didn’t help. Upscaling cleans up compression artifacts but doesn’t fix a spatial understanding problem.

The quality ceiling at this point was set entirely by the input images, even though I had HD source from multiple angles.

The Breakthrough: Press Photos Beat Video

About three experiments in, I tried a press photo instead of a video capture. Same play, same player — but a wire photo from a professional photographer. Sharp, no motion blur, visible floor and three-point line, clear spatial depth.

For AI input, the difference is dramatic. Court lines rendered with reasonable accuracy. Player proportions held. The AI generated angles from this image weren’t perfect, but they were coherent — they looked like plausible interpretations of a real moment.

This inverted my assumption. I had been thinking about this as a video problem. The best inputs aren’t video frames at all — they’re press photos.

A decades-old press photo can produce better results than 4K broadcast footage with motion blur. Resolution isn’t the variable. Clarity and spatial information are. For media companies, this reframes which archives are valuable. The photography archive — Getty, AP, wire photos going back decades — is potentially more useful for this workflow than the video archive.

The Slot Machine Problem

At some point in every session, the same frustration hit me: very limited control over what comes out.

Same image. Same parameters. Different random seed. Sometimes Qwen correctly renders the back of a defender’s head. More often it puts a face on the back of his head — the model is pattern-matching, not reasoning about geometry, so it defaults to “faces face forward.” When you rotate 180° you’re as likely to get a mirror of the original as a true reverse view.

I ran more iterations than I expected before finding anything usable. A few things shifted the odds in my favor, and they were worth learning:

Source image selection matters more than any parameter setting. Images where the player is at the apex of a jump shot — a held, readable pose — work far better than images where they’re in explosive motion. Qwen was trained primarily on product photography and synthetic 3D renders, so the cleaner and more static the spatial information, the better it performs. A dunk with defenders challenging the play is nearly impossible. A free throw at peak extension is much easier for AI to understand.

Running cheap image iterations before committing to video generation saved a lot of time and money. For video generation, I tested several models, and Kling V3 was the best. But it is an expensive step in the pipeline. Burning five Qwen attempts to find a generated frame worth interpolating costs pennies. Burning five Kling attempts on bad source frames costs around $3-4 per 10 seconds of video, and often that money produces nothing useful.

In this workflow, post-processing isn’t a final polish—it’s a core creative lever. Using motion blur and speed adjustments doesn’t just hide AI artifacts; it directs the viewer’s eye away from the synthetic and toward the action.

The Best Result: 360° Orbit Around Brunson’s Game-Winner

The final Brunson clip is a 360° orbit — the camera sweeps around his back, continues overhead, swings around to the front where you’re looking at him from the defender’s side, and returns to the source. It opens and closes on the real press photo: Brunson #11 in his release, Ausar Thompson #9 with his arm up, a referee in the background. Recognizable, real.

What works is the defender’s view: Thompson’s back, Brunson rising, ball frozen in mid-arc, and behind them the basketball hoop clearly visible, arena lights above the crowd. That frame looks like a real camera angle. It looks like something that should have existed and didn’t.

Is it clearly AI-generated? Yes — if you look closely the jersey numbers are wrong and the overhead rotation is obviously synthetic. Would it stop someone scrolling? Also yes. The technique is combining two Kling passes going opposite directions around the subject and trimming the junction so the camera moves through without dwelling on the distortions.

The Ewing Experiment: Swapping Avatar Characters

For the Patrick Ewing dunk over Alonzo Mourning in the 1997 playoffs, I replaced the players with AI-generated avatars styled after Space Jam. The Ewing stand-in I called Team Nighthawks — a comic-book basketball player with cornrows and motion lines around his feet. The Mourning stand-in became Team Terror Sharks — a green-skinned ogre with purple jersey and tusks. Characters were designed in Gemini Imagen, then composited into the broadcast footage using Kling’s video-to-video reference endpoint, which replaces visual elements while preserving the underlying motion of the original clip. The arena floor, the NBC crawl, NBA TV bug, the crowd, the original sound — all stays. Only the players are replaced.

For the first several seconds, it works. The Team Nighthawks character drives baseline fluidly through a real arena. The Team Terror Sharks ogre trails behind in a wide defensive stance. It looks like Space Jam with a bigger production budget.

Then Ewing rises to dunk. The AI, faced with two bodies in physical contact, does what these models do when they can’t understand something: it merges them. Jerseys bleed into each other — orange and navy and purple all at once. Limbs interlock in configurations belonging to neither character. By the apex there’s a single impossible entity wearing both uniforms simultaneously. The aftermath is genuinely funny — a full creature-feature close-up, tusks out, glowing red eyes.

I left all of it in. The merge, the chaos, the creature close-up. This is the clip.

The merge isn’t a bug to apologize for — it’s documentation of where the technology actually is. These models handle single subjects well and two-subject scenes acceptably, until the subjects overlap due to physical contact. At that point, the model has no coherent way to understand where one body ends and another begins, so it doesn’t try. The failure tells you something real: character substitution works in the right context — catch-and-shoot plays, free throws, celebrations — but a contested dunk is the worst possible test case. I chose it anyway, and the creature close-up at the end is more informative than any clean result would have been.

Given how quickly these models have improved over the past year, I’d expect these problems will fade. But right now, any production application for character substitution means you will probably get better results choosing plays without contact — and properly licensing the footage and player likenesses involved.

The Third Approach: Reconstructing What Was Always There

After the first session I went looking for a technique that didn’t rely on hallucination to create anything. That’s where Apple’s SHARP comes in.

The core difference from Qwen is philosophical. Qwen generates new pixels by predicting what a scene might look like from an uncaptured angle. SHARP infers the 3D geometry implied by the pixels that already exist, then lets you navigate that spatial geometry. It’s not making anything up — it’s reconstructing what the photograph encodes about 3D space. Same input always produces the same output — no slot machine — and it runs on Apple Silicon in a few minutes. Free.

The test image that produced my best result wasn’t Brunson in 2025. It was Walt Frazier shooting a jumper in Game 7 of the 1970 NBA Finals.

The source was a 1970 press photograph — grainy, slightly soft. I ran it through Real-ESRGAN (open source, free) to upscale and sharpen edges, then through SHARP. The whole local pipeline took under 20 minutes.

The output is a .ply splat file that loads in SuperSplat, a lightweight browser viewer. For this piece, the deliverable is a screen recording navigating within that viewer. In a production context, the viewer itself could be embedded directly in a web page.

The result: an immersive and navigable scene where Frazier occupies the foreground at jump shot apex, Jerry West contests with arm extended behind him, a referee stands at mid-court, Erickson watches from the right, his “24” jersey readable even from behind. The Madison Square Garden crowd fills the background. The court floor extends believably toward mid-court.

You can fly through it and orbit around it. The scene holds for roughly 30-40° of camera movement before edges start to tear — that’s empirical from this session. Within that range it looks geometrically correct, not hallucinated.

This is the archive equity argument made concrete. The Frazier splat works because it’s a courtside wide-angle photo with multiple subjects at distinct depths and visible court geometry — exactly the kind of shot photographers were taking in 1970. The broadcast footage from the same game is nearly unusable for AI tools like this. The archive that matters isn’t the video archive. It’s the photography archive.

The comparison between SHARP and Qwen/Kling comes down to a clear tradeoff: Qwen/Kling can orbit 360°, but jersey numbers are wrong, human anatomy can be visibly distorted, and anything outside the source frame is invented. SHARP doesn’t hallucinate — what you see is what was in the photo — but its range is limited and the scene tears at the edges.

One generates a convincing lie. The other reconstructs a constrained truth. For archive footage where fidelity to the original moment matters, the constrained truth is more interesting.

The open question for any archive application is whether this pipeline runs at scale across thousands of images. If it does, it changes how sports properties should think about cataloguing and investing in their photography archives — not just the video archive, but the decades of wire photos, press photography, and Sports Illustrated frames sitting in digital vaults. That’s worth investigating.

Update, March 12, 2026: After publishing this piece, I built the 3D viewer. You can explore the 1970 NBA Finals scene featuring Walt Frazier at viewer-henna-one.vercel.app — optimized for mobile, best on iPhone. Tilt your phone to orbit, drag to rotate. The source code is on GitHub. The whole build — project brief, iteration, deployment — took one additional session with Claude Code.

The Broader Point

The three approaches aren’t competing methods — they’re answers to different questions about what “remixing” sports footage might become in the age of AI.

Perspective synthesis asks: what was always there but never captured? The ceiling is photorealism. The failures are hallucinations and glitches in depth perception and player placement.

Character substitution asks: what new thing can this footage become? The ceiling is creative range. The failures are about model coherence when bodies interact.

3D reconstruction asks: what was encoded in this image that we haven’t been able to access? The ceiling is the fidelity of the original image. The failures are at the edges of the frame, where a single camera lens cannot see.

What I didn’t expect was how much of this turned out to be craft rather than configuration. Not a pipeline — a skilled human who can read a source image, knows which plays to try, understands when to lean into the distortion rather than fight it, and can tell the difference between a failed render and a weird asset worth keeping. That combination of creative judgment is what makes these projects come to life. These tools are still new and underused relative to what they’re already capable of.

That grainy 1994 clip that stops the scroll for nobody. A single wire photo from that season — sitting in a Getty vault right now — could be an immersive 3D scene in twenty minutes of local compute. Nobody’s doing this yet at scale. That gap won’t stay open much longer.

Leave a comment

Not a Substack post. Not a data journalism piece. A real-time labor market tracking tool — something like a Bloomberg Terminal for knowledge worker displacement signals, with live data feeds, a proprietary scoring system, and color-coded severity indicators. I had a working prototype with eight signals, a composite index I called the Knowledge Worker Displacement Index, and an actual deployment pipeline.

I killed it on day two. Here’s why — and what that decision taught me about the difference between building something impressive and building something useful.

Where It Started: A Hypothesis Looking for Evidence

I’m job searching right now — unexpectedly, after a long run at NBC News Group. And I started noticing a pattern I couldn’t explain away — not in the data, but in the conversations I was having with former colleagues, people in my network, peers who’d been laid off from knowledge worker roles. Senior people. Experienced people. Seven months into searches with nothing to show for it. Meanwhile the headline numbers looked fine. Unemployment at 4.3%. S&P near record highs.

I’d spent almost a decade leading product teams and working alongside editorial newsrooms building apps and content platforms for NBC News, MSNBC, and CNBC, watching AI reshape media from the inside. I had a hypothesis: something structural was happening in the knowledge worker labor market that the standard metrics weren’t designed to see. But a hypothesis isn’t a story. I needed to find out if the data agreed.

So I started building.

Day One: The Dashboard

The first version was a React web application — a live tracking tool pulling from BLS, LinkedIn Economic Graph, Challenger Gray & Christmas, and the NY Fed. I built a composite scoring system that weighted each signal by severity and rolled them up into a single index. It looked like something between a financial terminal and an election forecast model.

I was proud of it. It was technically ambitious. It updated with real data. It had a color-coded alert system — critical, elevated, normal — that gave each signal a status based on how far it had moved from historical baselines.

Then I did something I’ve learned to do with every product I build: I asked someone to tear it apart.

The Adversarial Critique That Changed Everything

I used Claude to evaluate the dashboard — not as a collaborator helping me improve it, but as a skeptic. I prompted it to argue against what I’d built from multiple angles: a data journalist, a policy wonk, a hiring manager looking for reasons to dismiss it. And critically — a financial analyst.

That last persona produced the feedback that stung the most — and mattered the most.

Here’s what I hadn’t fully admitted to myself: the dashboard wasn’t just trying to track knowledge worker displacement. It was trying to make a second, larger argument — that the divergence between financial market performance and white-collar hiring conditions represented an underpriced risk. A “Big Short” style thesis. Equity markets were near record highs while the professional labor market was quietly deteriorating. I thought those two things in combination were a signal worth building around.

The financial analyst critique was unsparing: the data I had could support the labor market story, but it couldn’t support the financial prediction. The causal chain between displacement signals and market mispricing was too long, too confounded, and too dependent on assumptions I couldn’t defend. A financial analyst reading this wouldn’t find it persuasive — they’d find it a stretch.

Once I saw that clearly, the path forward wasn’t to find a more forgiving format — it was to make a more defensible argument. The labor market story was real and the data supported it. The financial trade thesis was a reach, and stapling it to the labor market analysis was making both weaker. The composite index had the same problem: it was doing work the data couldn’t support — rolling eight signals with different methodologies and different underlying populations into a single authoritative number implied a precision that wasn’t there. Stripping out the financial prediction angle and focusing on what the evidence actually showed wasn’t lowering the bar. It was finally clearing it.

I’ve started treating adversarial prompting as a structured part of the build process — not a last check before publishing, but something I run early enough to change direction. It’s also one of the most useful things I’ve found to do with a product team before a major decision: not “what do we think about this?” but “make the strongest case against it.” The goal isn’t to find reasons to kill what you’ve made. It’s to find out if it can survive against counter-arguments seeking to kill it.

Day Two: Building the Charts as Editorial Decisions

I still thought of this as a standalone editorial product — a custom site, interactive charts, something I could build, deploy, and share as its own thing. Then I asked the same question I’d asked about the dashboard: who is this actually for, and what’s the simplest surface that serves them? The answer was Substack. The argument didn’t need its own infrastructure. It needed to reach the right readers.

That decision changed how I thought about every visualization. Interactive charts became image assets — which forced a useful constraint. You can’t hide behind interactivity when a static image has to do all the work on its own. I used Claude Code to write an export script that pulled cleanly formatted charts and signal cards directly from the tools I’d built, with sourcing and contextual commentary baked in. The tool handled the formatting. Every decision about what to show, what to cut, and what the chart needed to say — that was the work the tool couldn’t do. The dashboard never shipped. But the work inside it did.

The clearest example: the S&P vs. White-Collar Hiring Rate chart.

I had half a dozen ways I could have framed that data — indexed to a shared baseline, shown as separate panels, broken out by sector. I chose two lines on the same axis, diverging since early 2023, with no embellishment. Not because it was the most technically interesting option, but because it was the most direct one. If a reader bounced after the first scroll, I wanted them to have already seen the argument. Everything else in the piece is evidence for what that image says.

The signal cards from the dashboard survived, but transformed. Instead of a live tracking UI, they became a structured appendix — a reference section with sourcing, severity ratings, and historical context. The critical/elevated/normal system I’d built for the dashboard turned out to be genuinely useful as editorial framing, even after the composite index it originally fed was gone.

Hypothesis to published article: two days.

On Building as Thinking

My background is product management. I spent nearly a decade leading digital product for news publishing and streaming platforms — working closely with editorial teams, data journalists, video producers, and graphics editors who do this kind of work professionally and under real deadline pressure. I understand what rigor looks like in a newsroom context, which is part of why the adversarial critique mattered: I knew what a skeptical editor would say before they said it.

What I brought to this project wasn’t engineering expertise. It was product judgment: knowing what to build, for whom, and when to stop. That combination — product thinking plus AI-augmented building capability — is what made a two-day turnaround possible. And I think it’s increasingly what effective product leadership in consumer media requires. Editorial organizations are building AI-assisted products faster than most product teams know how to evaluate them. You can’t lead teams well if you can’t also do the building — not because you need to ship the code yourself, but because you need to know where the real decisions are being made and whether they’re being made well.

What I’ve come to believe — and this felt like a personal realization more than a general principle — is that the process I used here isn’t really a product process. It’s the editorial process. Hypothesis, evidence gathering, structured self-critique, pivot, rebuild, publish. It’s what good data journalists and researchers have always done, just radically compressed by new tools. The judgment required — what to build, who it’s for, when the evidence actually supports the claim — those aren’t new skills. They’re the skills that newsrooms have been developing for a long time. Product teams are just now learning to use them.

I built this piece because the story felt personal and the data felt real. The job search that prompted the hypothesis also gave me the time and motivation to see it through — which isn’t nothing. The tools made a two-day turnaround possible. But the decisions that made it worth reading were the ones that required killing something I was proud of, admitting an argument I couldn’t defend, and choosing the surface that served the reader over the one that looked most impressive.

Read the original data-driven piece this process produced here:

Leave a comment

On the surface, the numbers support him. The S&P is near record highs. Headline unemployment is 4.3%.

But a few days before that speech, the Bureau of Labor Statistics quietly released revised data showing that the U.S. economy added just 181,000 jobs in all of 2025 — revised down from an initial estimate of 584,000, and the worst year for hiring since 2003 outside of a recession. The labor market actually contracted in four separate months last year. January 2026 added 130,000 jobs, almost entirely in health care and construction — not the knowledge work roles where the labor market is effectively frozen.

I spent years at NBC News Group at the intersection of AI, media, and technology — watching AI reshape industries from the inside. What I keep seeing in the labor data doesn’t match what was said in the Capitol last week. The metrics the president cited were designed to measure a different kind of economic problem than the one unfolding right now.

The unemployment rate measures people who lost jobs. It doesn’t measure people who can’t get them — the engineer in month seven of a search, the graduate with 200 applications and five callbacks. The signals that capture that reality look very different from the headline. And the divergence between the two has been widening every quarter.

Something is happening inside the knowledge worker labor market that the aggregate statistics aren’t designed to see. This is what it looks like in data.

The stock market and knowledge worker hiring have diverged sharply since early 2023 — a 110-point gap that has widened every quarter. This divergence isn’t coincidental: equity markets are rewarding companies for reducing headcount, which means the same forces making investors richer are making knowledge workers’ job searches harder.

The Freeze

The labor market for knowledge workers isn’t slow. It’s frozen.

This is happening two ways simultaneously: existing roles being eliminated, and new roles not being created. The first shows up in layoff announcements. The second is harder to see — it lives in hiring rate data that most people never look at.

Job openings have fallen 46% from their 2022 peak. But the more telling signal is the hires rate — actual job starts, not posted openings. Hires are at 2013 levels. Even where openings nominally exist, companies are running interview processes they’re not closing. The labor market is frozen at both ends: cuts happening, hiring not following.

The gap between openings and actual hires is widening — companies post roles but don’t fill them. January 2026 layoff announcements hit 108,435, the highest since 2009. AI-cited cuts are growing as a share of total.

Structural Signals

The canary is recent college graduates.

The clearest signal that this is structural, not cyclical: recent college graduates are now more likely to be unemployed than the overall population — a historic reversal. New entrants to the labor market take whatever the market offers. When they can’t find knowledge worker roles, it means those roles aren’t there, not that conditions are temporarily tough.

Underemployment for recent graduates hit 42.5% in Q4 2025 — its highest level since 2020 — meaning nearly half of new graduates who found jobs are working in roles that don’t require a degree. The bottom rung of the knowledge worker career ladder is being removed.

J. Scott Davis, an economist at the Federal Reserve Bank of Dallas, argues the disruption runs deeper than a market correction. The current knowledge-worker career model depends on entry-level roles where new hires perform codifiable tasks while building tacit knowledge over time. Davis writes that AI is making this method of employee development 'cost-ineffective' for firms1 — removing the bottom rung of the career ladder before the next generation has a chance to climb it.

If entry-level disappears, the talent pipeline breaks. There will eventually be a senior talent shortage — but only after a prolonged period of displacement for the cohort entering the workforce now.

The sharpest confirmation of the AI-displacement thesis comes from job postings data. Indeed Hiring Lab segments postings by AI exposure — roles where AI can perform a significant share of the tasks versus roles where it can’t. The divergence is unambiguous.

This chart answers the counterargument directly. If this were a broad post-ZIRP correction, you’d expect all job postings to contract together. Non-AI-exposed roles are flat. The decline is isolated to the category of work that AI can perform. That’s not a cyclical pattern. That’s a structural one.

The strongest counterargument

The most credible alternative reading: this is a post-ZIRP correction, not a structural AI transition. Companies massively over-hired in 2020–2022 when capital was free and remote work opened every geography. They are now normalizing to sustainable headcount levels — and AI is convenient cover for cuts that were coming anyway. Several economists hold this view seriously, and some of the layoff data is consistent with it.

The case against the correction-only explanation: three years in, a true post-pandemic correction should be finishing, not accelerating. The companies cutting hardest — Block, Pinterest, Salesforce — are profitable and growing, not distressed. Jassy told Amazon employees in June 2025 that AI would reduce white-collar headcount “in the coming years” as a matter of strategy, not correction. And critically, the hiring freeze is spreading to companies that never over-hired at all. A correction explains some of what’s happening. It doesn’t explain all of it.

The Incentive Structure

The stock market is rewarding this.

Block is distinct from other recent layoffs. Amazon's 30,000 cuts were framed as "removing layers" and "de-bureaucratization" — restructuring language without a direct AI capability claim. Pinterest explicitly attributed cuts to AI reallocation. Workday named AI in its restructuring. But Block is the first S&P company to make a 40%+ cut while explicitly crediting AI capability as the primary mechanism — and to predict that most companies would follow within a year.

This creates a compounding incentive. Every positive market reaction to an AI-attributed cut makes it easier for the next CEO to justify the same move to their board. The mechanism isn't just that AI enables cuts — it's that equity markets are now explicitly pricing in the expectation that companies will pursue them.

The Policy Trap

The standard policy responses don’t fit this problem.

When the labor market deteriorates, the Fed cuts rates and Congress passes stimulus. Both mechanisms were designed for cyclical recessions, where the problem is insufficient demand. This displacement is different in kind, and the standard toolkit is partially mismatched.

Rate Cuts Won’t Restore Hiring

If a company eliminated software roles because AI performs that work, borrowing cost is not the constraint on rehiring. Whether rates are 5% or 2% doesn’t change whether a company that just achieved 40% efficiency gains rehires into eliminated roles. What rate cuts will do: reflate asset prices, helping people who own assets — disproportionately not the displaced knowledge workers described here.

Fiscal Stimulus Hits an Inflationary Environment

Direct payments support spending but risk adding pressure to an environment that hasn’t returned to pre-pandemic price levels. A specific feedback loop: rate cuts could reinflate housing costs in the high-cost metros where this cohort lives — making their affordability problem worse at the exact moment they can least afford it.

Retraining Has a Poor Track Record Here

Retraining works when there’s a clear adjacent pathway and the destination occupation is growing. As Davis argues: if AI is making the entry-level development model cost-ineffective, you can redistribute who gets the remaining jobs. You cannot train your way to more jobs than the economy needs.

Unemployment Insurance Was Designed for a 26-Week Gap

New York raised its max weekly benefit to $869 in October 2025 — still just 24% income replacement for a $150K earner. California’s cap remains at $450/week (unchanged since 2005), a 15% replacement rate in one of the world’s most expensive labor markets. These are short bridges for workers who find equivalent work quickly — not for displaced knowledge workers facing a structurally altered market. If claims spike, state trust funds are the first systems to strain, triggering automatic employer tax increases that further dampen hiring.

This is not an argument for doing nothing — inaction would be worse. It is an argument that the transition unfolding here was not anticipated by the policy infrastructure designed to manage labor market disruption. The instruments are blunt and affect the whole economy, not specific labor market segments.

The aggregate labor market numbers will catch up eventually. They always do — with a lag, after the people living it have already felt it for a year or two.

What happens in that gap matters. Whether this is a correction finishing or a structural shift accelerating will become clear in the data over the next few quarters. But for the cohort entering the knowledge worker labor market right now, the distinction may be academic. The freeze is real either way.

If something here is wrong or missing, I’d like to know. Leave a comment or find me on LinkedIn.

Leave a comment

Signals To Watch

4 critical · 7 elevated · 1 normal

Hiring Freeze

Structural Displacement

Fiscal Exposure

Methodology & Limitations

The signals above are relevant to the knowledge worker displacement thesis — the argument that AI is driving a structural reorganization of white-collar employment that aggregate unemployment statistics are not designed to measure. Signals are drawn from public sources and represent a snapshot as of February 27, 2026. This is not financial advice. The thesis presented here is speculative and contested; see the Yale Budget Lab, Dallas Fed, and Stanford Digital Economy Lab for the academic evidence landscape.

I’m currently looking for my next VP Product role. If you’re building AI products and care about responsible design, let’s connect.

Try the tool: NYC School Data Explorer · View the code: GitHub

Last month I got a text notification from the NYC Department of Education asking me to fill out a school survey. As a parent with two kids in public schools, I tapped through and filled out the survey for their school. As a product builder, I started wondering: what actually happens to that data?

It turns out those surveys feed into something called School Quality Reports — massive datasets covering over 1,800 schools, measuring everything from test scores and student growth to parent satisfaction and teacher retention. The data is publicly available. It’s also dense, hard to navigate, and easy to misinterpret.

So I built an AI-powered tool that makes it accessible to education reporters. Along the way, I learned more about responsible AI design than I expected — and confirmed something I already suspected: strong product skills transfer across domains.

The Problem with School Rankings

Anyone who’s looked at school data knows the temptation: sort by test scores, pick the “best” ones, done. The problem is that high test scores often tell you more about the median income of the families of enrolled students than about what’s actually happening inside the classroom.

The NYC School Quality Reports provide two different lenses on performance. There’s a Performance Score — absolute achievement, basically raw test results and graduation rates. And there’s an Impact Score, which measures student growth: how much progress students make relative to similar peers citywide. NYC recently introduced the Impact Scores, and the two most recent academic years include this metric.

When I ran the correlations, the difference was stark. Performance Score correlated with poverty at -0.62. Impact Score? Only -0.28. That gap is enormous. It means Performance Score is heavily confounded by wealth — wealthier student bodies score higher almost by default. Impact Score strips a lot of that out. It’s a much better signal for what a school actually contributes to student learning.

This is where it got interesting. I started looking at high-poverty elementary and middle schools — those where 85% or more of students qualify as economically disadvantaged — and applied a growth-based framework to see which ones were actually moving the needle on student learning. Then I checked whether the pattern held up over time. It did: 100 high-poverty schools maintained strong growth status across both 2023-24 and 2024-25. Same schools, same exceptional growth, two years running.

Within that group, the most compelling stories are what I started calling “Hidden Gems” — schools achieving exceptional growth despite lower absolute test scores. Traditional rankings would bury these schools because their raw numbers don’t look impressive. But their students are growing faster than peers citywide. That’s the pattern worth investigating.

Using AI for Product Research

Before writing any code, I needed to understand the data landscape. This is where I used AI in a way that might surprise people.

I asked Claude to help me search NYC’s open data ecosystem for complementary datasets — school budgets, PTA finances, suspension records — that could add context to the quality reports. Claude found datasets published on NYC’s InfoHub site that weren’t surfaced through the main Open Data portal. It navigated between multiple government data repositories, figured out what was available and what format it came in, and saved me what would have been days of clicking through government websites.

One of my beta testers, an engineering manager at an NYC education technology provider, noticed this capability and commented: “The way it’s able to navigate frequently clunky government websites is itself very interesting.” That observation captured something people miss about AI in product development. Using AI for discovery and research — before you build anything — is a product skill, not just a technical one.

Working with Journalists

I didn’t build this based on assumptions about what people might want. I reached out directly to education journalists — people I’d never met — and asked: what questions do you need answered? What makes the current data hard to work with?

I wasn’t sure she’d respond, but I cold-emailed the NYC Bureau Chief at Chalkbeat — one of the most respected education newsrooms in the country. She agreed to beta test an early version, and her feedback reshaped the tool. She needed sorting by different metrics, attendance data alongside test scores, teacher experience information to contextualize performance. Working with Claude Code, I shipped those changes in one day and quickly brought those updates to her attention.

Her response: “This tool is fascinating and extremely helpful.” She started sharing it with her colleagues and specifically noted the features I’d added since her first session — attendance data, teacher turnover — which told me she was actually using it, not just being polite.

You don’t assume you know what users want. You ask, you ship, you learn, you adjust. The difference when building with AI is that this loop accelerates, so you can get feedback and learnings faster, thus product decision making becomes even more critical.

Responsible Design: What to Show, What to Hide

The most important product decisions I made weren’t about features. They were about constraints.

When you build an AI interface for data that affects where families send their children, design choices become ethical choices. I made four that shaped everything else.

The tool doesn’t rank schools. It would be trivially easy to sort schools from “best” to “worst” and return a ranked list. I chose not to. High-stakes decisions about children’s education need human judgment, not algorithmic sorting. The tool surfaces data and context; journalists and families do the analysis.

The tool highlights schools beating the odds, not just high performers. Traditional rankings favor wealthy neighborhoods by design. By focusing on schools achieving exceptional growth in high-poverty communities, the tool surfaces stories that traditional approaches miss entirely. These aren’t presented as “the best schools” — they’re presented as patterns worth investigating.

Every response includes the Economic Need Index alongside performance data. You can’t interpret a school’s test scores without understanding who it serves. Stripping socioeconomic context from education data isn’t just misleading — it can reinforce the same segregation patterns the data should help us understand.

The tool raises questions rather than claiming answers. I can tell you that 100 high-poverty schools produce exceptional student growth across two consecutive years. I can’t tell you why. That requires visiting schools, interviewing principals, understanding culture and leadership on the ground. The tool is a starting point for reporting, not a final word.

Building Quality Systems

You can’t ship AI products without measuring quality. At NBC, we monitored API error rates, page load times, and video streaming health. For an AI tool, you need to monitor the quality of the outputs themselves.

I built an evaluation system that scores every AI response on a 100-point rubric across five dimensions: factual accuracy, context inclusion, limitation acknowledgment, responsible framing, and query relevance. Responses scoring below 75 get automatically logged to a Google Sheet via Zapier, where I periodically analyze patterns with Claude Code and adjust system prompts.

One of the more interesting architectural decisions was using different AI models for different jobs. Claude handles the complex reasoning and response generation — it’s better at nuanced analysis and following detailed instructions about responsible framing. But for evaluation scoring and the contextual follow-up suggestions the tool offers after each response, I use Gemini Flash 3. Evaluation and suggestion generation need to be fast and cheap, not deeply analytical — they’re high-volume tasks where speed and cost matter more than sophisticated reasoning. The cost difference is real: I’ve spent roughly $30 on Claude API credits building and running this tool, and about $1 on Gemini. That’s a 30:1 ratio — and it matters, because if you’re evaluating every single response, the evaluation layer can’t cost more than the response itself. Keeping the evaluator independent from the system it’s evaluating is also just good practice; you don’t want a model grading its own homework.

This kind of model selection — matching the right tool to the right task based on capability, cost, and architectural constraints — is a product decision, not just an engineering one. It’s the same thinking I’d apply to any system at scale: what’s the unit economics story, and does each component justify its cost?

Within days of deploying the evaluation pipeline, it caught a factual accuracy issue that I fixed by adjusting the system prompt. Hours to identify and resolve, not weeks. Same principle as monitoring any production system — you measure what matters and build feedback loops. The difference is that with AI, “what matters” includes whether the outputs are actually correct.

Technical Execution

The tool is built on MCP — Model Context Protocol — which is an open standard for giving AI systems structured access to external data. I designed a set of MCP tools for NYC school data: search, school profiles, comparisons, correlations, and chart generation. Then I built it with Claude Code and deployed a Next.js frontend on Google Cloud Run.

Here’s what makes MCP interesting as an architecture choice: the website is just one client. The same MCP server and tools can plug directly into Claude, or any AI assistant that supports the protocol. I was fact-checking this article’s claim about those 100 schools by querying my own MCP tools live inside Claude. Same data, same tools, different interface. That’s the power of building on an open standard rather than a closed application: the data layer is reusable across any AI context, not locked to a single frontend.

The responsible design principles aren’t bolted on after the fact. There’s a three-layer system: a pre-filter that catches harmful query patterns before they reach the AI, a system prompt that guides reasoning and requires context in every response, and tool-level constraints that force the inclusion of economic need data alongside any performance metric. If someone asks the tool to rank schools or filter by demographics, it doesn’t comply — it explains why and offers a more useful alternative.

You can try the tool at nyc-school-explorer-w27tadi35a-uc.a.run.app or explore the code on GitHub.

Questions, Not Answers

Some elementary and middle schools in this data show outlier growth despite serving populations with the highest economic need in the city. One hundred of them do it consistently across both years we have data for. Their attendance rates, instructional quality ratings, family trust scores, and teacher retention all point in the same direction: something effective is happening at these schools.

But understanding why is harder, and more important. Is it leadership? Culture? Specific instructional practices? Community partnerships? The data can identify the pattern. Journalists and researchers have to investigate and explain it.

That’s a feature, not a limitation. Responsible AI means knowing what AI can’t do.

Who’s Using It

Chalkbeat’s NYC Bureau Chief beta tested the tool and shared it with her colleagues. Education technology professionals have been testing and asking sophisticated questions about the methodology. NYC Department of Education staff are sharing it internally. The Manhattan Borough President’s office has expressed interest.

These aren’t hypothetical users. They’re people who work with education data every day, and they’re finding it useful. That’s the validation that matters to me as a product builder — not download counts or demo impressions, but whether the people it was built for actually reach for it when they’re doing their work.

What’s Next

I’ve been building media products for twenty years. I built this civic data tool for education reporters in a couple weeks. The core product skills — customer engagement, data fluency, responsible design, iterating on real feedback — are the same. The domain is just context.

The NYC School Data Explorer is live for education reporters during survey season, when parents across the city are evaluating their schools. I’m already planning my next project exploring different AI capabilities, because the best way to demonstrate product range is to keep building across different domains.

If you’re hiring for VP Product and want someone who builds AI products responsibly, let’s connect.

Try the tool · View the code · Connect on LinkedIn

Leave a comment