You interrupt your own flow to check /context, racing to save your work before compaction eats it. That vigilance is a tax — and you can automate it away.
Generating technical documentation from a codebase is easy now. Trusting it isn't. Evidence-linked claims and human sign-off are the missing steps.
Comprehension debt finally has a name. The popular cure — write more docs — inherits the gap it's meant to close. What actually pays it down.
Compaction keeps a coding agent running by quietly deleting the why, so you become its memory. What gets lost, why it compounds, and the local-first fix.
LangChain OpenWiki validates repo documentation for coding agents. But repo orientation is not the same as evidence-graded product understanding.
Your vibe-coded app caught on and now people pay for it — but you can't fully explain it. Build a mental model without reading every line, top-down.
Your product works. On launch day someone asks what it does at the edges — and you'd have to open the file to answer. That gap is the real launch risk.
Vibe-coding got you a working product you can't speak for. Reclaiming it isn't slowing down — it's re-authoring what you shipped, behavior by behavior.
You own a product you can't explain — acquired, inherited, or vibe-coded with AI. Here's how to read code you didn't really write, behavior by behavior.
You vibe-coded a working product. Users use it. You can't speak for it without reading every line. The product is yours — you just don't know it yet.
A spec extracted from code reads what is there, not what was decided — so placeholders and fallbacks get promoted to intent. Sign-off is the missing step.
Execution memory has tooling. Decision memory is emerging. The behavior layer — what your product promises to do — is the one almost nobody has built yet.
AI made the rewrite cheap; recreating years of product decisions from memory did not get cheaper. How to keep the decisions when you throw away the code.
Paste a prompt into Claude or ChatGPT, describe your product module briefly, and get a .pbc.md behavior spec you can view, edit, and commit to your repo.
CLAUDE.md and AGENTS.md tell agents how to work in your repo. They don't tell agents what your product promises. That's a different artifact — the PBC layer.
A step-by-step guide to writing a .pbc.md file for your product's most critical module. Start with plain Markdown; add structured blocks agents can read.
Shipping fast with AI agents feels productive. But the costliest mistake isn't bad code — it's building confidently when nobody wrote down what was decided.
PRDs capture intent. Tests verify assertions. Between them, there's no artifact tracking what the product promises — grounded in code, confirmed by humans.
AI agents have AGENTS.md, memory banks, harnesses, and monitors. They still lack the product context layer — what the product promises and what must hold.
AI can extract product logic from your codebase. Stewie builds a living behavior spec your whole team can read — no code, no docs, no waiting on engineers.
Your agent ships code that passes review and still breaks a product promise. It is not a capability gap — it is the product context no rules file gives it.
Your repo has workflow instructions, session context, and feature specs. None of them answer what the product promises to do. That's the PBC layer.
Shipping fast with AI coding tools is genuinely good. The problem isn't the speed — it's what gets left behind. Product reasoning doesn't survive the vibe.
