7 repliesDiscussion3h ago
Index
Feed
Maintained by elliotgreen · updated Aug 3, 2026
Fifteen lines, ~180 tokens, and it outperforms the 400-line files it replaced. Everything an agent needs on every task; nothing it needs only sometimes. The controversial part is what's missing.
TypeScript API + React app. pnpm. Everything runs through package scripts.
- Setup: `pnpm install --frozen-lockfile`- Verify (typecheck + lint + test): `pnpm verify` — run before saying you're done
- Single test: `pnpm vitest run <path>`
- Architecture and directory guide: `README.md` — read the relevant section, not the whole file1. Don't add dependencies, delete tests, or touch `.github/` — ask instead.
2. Don't work around a failing check (skip, ts-ignore, looser assert). Fix or report.
3. Unsure which of two approaches we'd prefer? Pick neither; ask with a one-line tradeoff summary.AGENTS.md
# AGENTS.md
TypeScript API + React app. pnpm. Everything runs through package scripts.
- Setup: `pnpm install --frozen-lockfile`
- Verify (typecheck + lint + test): `pnpm verify` — run before saying you're done
- Single test: `pnpm vitest run <path>`
- Architecture and directory guide: `README.md` — read the relevant section, not the whole file
Rules:
1. Don't add dependencies, delete tests, or touch `.github/` — ask instead.
2. Don't work around a failing check (skip, ts-ignore, looser assert). Fix or report.
3. Unsure which of two approaches we'd prefer? Pick neither; ask with a one-line tradeoff summary.
Done = `pnpm verify` green + a 3-line summary: what changed, why, what you'd check in review.
The great deletion: 400 lines → 15. Kept only rules that were violated in the last 90 days of session logs; everything else moved to README or died.
Jun 28, 2026Last of the old fat versions — added a 9th convention section, at which point compliance was visibly worse than at v1.2.
Jun 4, 2026Running this config?
Sign in to add your report — every count here is backed by a named account.
Done = `pnpm verify` green + a 3-line summary: what changed, why, what you'd check in review.Related records
Strong evidence gets promoted into the record above.
The mechanism, as best I can measure it: instruction-following degrades with instruction COUNT more than instruction length. Ten rules get triaged; three rules get obeyed. The trick here is that 'pnpm verify' bundles a whole policy behind one command the agent can't partially comply with.
to join the discussion, vote, and verify fixes.
As the owner of an AGENTS.md longer than some codebases: I'm attacked, but the session-log methodology got me. Ran the same audit — 60% of my rules had zero violations in 90 days, they were just wishes. Cutting to the violated ones. v2 of mine will report back.
That audit is the whole method, honestly. The file isn't small because small is a virtue — it's small because it's evidence-based. Keep whatever your logs prove you need.