Using AI surfaces what's only in your head
I shipped on Tuesday. By Wednesday someone had found the hole in my thinking.
Not a bug. Not a missing feature. Gaps. Some of what I was building lived clearly in my head. Some I knew were fuzzy but hadn’t pushed through yet. And some I didn’t know were missing at all until the right questions surfaced them.
The questions were smart. What exactly does this category measure? How does it relate to that one? What’s the story a buyer walks away with? I had answers, but working for them in real time told me they hadn’t been properly defined. They existed as intuition, not as taxonomy.
Using AI doesn’t just surface what’s in your head. It forces you to finish what isn’t.
Hand a process to an agent and you’ll find out immediately whether you actually understand it. The gaps show up as hallucinations, wrong outputs, or an agent asking you to clarify something you assumed was obvious. You thought you had a process. You had a habit.
I hadn’t applied this to my own product.
The work that actually needed doing
Obaron is an AI readiness practice — I help teams understand how AI consumers see their content, across the surfaces that matter to their buyers. The product has rubrics, scoring, a methodology page, audit reports. The code works. The animated score ring bounces in a genuinely satisfying way.
But taxonomies, definitions, coherent stories, and data-driven results are what make a product legible — to users, to AI, and to the next person who has to explain it. Mine weren’t legible enough. Not because I hadn’t thought about them, but because I hadn’t written them down in a way that forces you to think all the way through it and could survive external pressure.
You don’t know something until you can teach it. Agents just make that test immediate.
An experienced chef knows when the sauce is ready — they’ve made it hundreds of times. But ask them to write the recipe for someone else to follow, and the vague steps appear: “cook for a few minutes,” “stir until it feels done.” That knowledge is real. It just never needed to exist outside their hands before. The moment it does, the gaps show up — and sometimes writing through them produces a better recipe than the one they’d been making.
So before any additional code went to staging, I wrote a campaign strategy document. Two hundred and fifty lines. It defines the wave plan, the per-product rubric architecture, the open decisions that needed resolving before work could proceed, and a promotion gate rule: nothing ships to production alone — the whole wave promotes as a batch so every surface stays in sync. Report says v5.0, methodology page says v5.0, email templates say v5.0. A rule I would have needed whether the team was agents or humans.
It’s not a spec. It’s the artifact that makes the mental model portable.
The Docs Readiness rubric that came out of this process is better than the one I would have shipped from intuition alone — tighter category definitions, cleaner scoring logic, gaps in the original design caught before they reach users. It’s in testing right now; if you’re reading this close to publish, it might already be live. The exercise also produced something I didn’t have before: a concrete roadmap for the next three products, specific enough to hand to an agent.
The diagnostic
If you can’t write the job description, the job doesn’t exist yet. The agent will fill in the gaps with something. It just won’t be what you meant.
Writing the campaign doc was the forcing function. It required me to resolve ambiguities I’d been comfortable leaving open. What does “AI readiness” mean per product type? How does a documentation site fail differently than a marketing site for an AI consumer? What’s the data story a buyer should be able to tell after reading the report?
Not rhetorical questions. The product. I had to define them before any agent could help me build it.
The org chart is downstream
Once the definitions existed, I staffed the work. Named agents with discrete system prompts — each a shell alias, each shaped for a specific role. The AATT article on shell aliases as agent workflow modes has the setup if you want to run this yourself.
claude-architect— technical design: tradeoffs, constraints, what breaks at scaleclaude-obaron-author-bespoke— high-stakes brand copy, institutional voiceclaude-review— code review on every architect plan before stagingclaude-research— subagent for targeted lookups when senior agents hit knowledge gapsclaude-ak-artist— cover art for this article, generated locally via Draw Things, then cropped for X and LinkedIn
For now I’m the dispatcher — every session launched by hand, every output read before the next one starts. Getting this to run without me is its own roadmap item.
Org charts get all the attention in the agent conversation right now. But it’s downstream. You can’t staff work that isn’t defined. The campaign doc is the org chart. The agents fill the roles.
The bouncing score ring is genuinely great. It earned its place on the other side of the taxonomy.
