sdlc-feedback-looplisted

# Let the agent check itself The highest-leverage thing in this whole set. Without it, the first reliable signal that code works arrives from CI, a reviewer, or production — and when an agent is producing the code, a late signal means a person verifies *everything*. That person is now the constraint, and adding agents makes it worse rather than better. With it, the session finds and fixes its own mistakes, and what reaches a human has already cleared the bar. ## Four moves **One command.** If verifying today means running four things and knowing which failures matter, wrap it: `bun run check`, `make verify`. It must exit non-zero on failure — an agent reads the exit code before it reads your prose. **Show what healthy looks like.** List the commands in `CLAUDE.md` with a line of expected output, so pass and fail are distinguishable without asking. **Give each task a target the agent can evaluate.** "The three cases in `persistence.test.ts` pass." "The endpoint returns 409 on a duplicate." "The rendering matches the mock at 375px." Not "make it work" — that has no closing condition, so the agent guesses at done. **Put verification inside the definition of done**, in `CLAUDE.md`: ```markdown ## Before you say it's done Run `bun run check` and `bun test`, and paste the output. Both must be clean — a skipped test is a failing test. If something fails, fix the code. Do not edit the test to make it pass, and do not delete it. If the test itself is wrong, say so and stop. ``