How this repo is built: I set the boundaries, agents write the code, tests decide what is true

[daichunghy]

I do not write the TypeScript in this repository myself. Most of it was produced by AI coding agents, and I could not have produced it without them. That is the working method, so here is the honest version, because the interesting part is not the code — it is what keeps the code honest.

What I do:

• Write the requirement as a contract: what must be true, and what is explicitly out of scope. The "not a fit" and "what is not shipped" lines in the READMEs come from me, not from an agent.
• Keep the boundary claims honest: no invented stars, no "production ready", no pilots that did not happen. Every status line references something you can check — a release tag, a CI run, a receipt.
• Review by evidence: CI, deterministic receipts, conformance fixtures. When a change breaks one, I do not adjudicate between me and the agent. The test does.

What the agents do: implementation, refactors, fixture generation, first drafts of docs.

What decides: npm run verify. Green means the change is real. Red means the agent is wrong, politely.

The failure mode I watch closest is an agent writing a confident README claim the code does not back. That is why this repo documents what it does not do in the same breath as what it does.

If you maintain something built the same way: what is the one check you would not delegate to the agent? That is the layer I am trying to describe, and I collect examples.