Prompts — a copy-paste loop library

# Prompts — a copy-paste loop library

Real loop prompts harvested from the wild, organized so you can grab one and
adapt it. **Every prompt is attributed to its source** — follow the links to see
it in context and to credit the authors who figured these out.

> How to use this: pick the file that matches what you're building, copy the
> prompt, and change the project-specific bits (paths, check commands, scope
> rules). Read the inline notes — the *guardrails* in these prompts (anti-cheat,
> scope limits, escalation) are where most of the hard-won lessons live.

| File | What's inside |
|------|---------------|
| **[ralph-loops.md](ralph-loops.md)** | The `while :; do cat PROMPT.md \| agent; done` technique and its refined `PROMPT.md` build/plan files, with the famous anti-placeholder "signs." |
| **[goal-and-loop.md](goal-and-loop.md)** | Claude Code `/goal` (stop-on-condition) and `/loop` (cadence) prompts, from report-only to auto-fix. |
| **[automations.md](automations.md)** | OpenAI Codex scheduled-automation prompts: daily triage, issue triage, PR babysitting, CI repair. |
| **[subagents-maker-checker.md](subagents-maker-checker.md)** | Explorer / implementer / independent-verifier sub-agent skills — the maker/checker split as prompts. |
| **[verification-rubrics.md](verification-rubrics.md)** | LLM-as-judge grader prompts that return PASS/FAIL with evidence. |
| **[loop-contracts.md](loop-contracts.md)** | Full, machine-readable loop specs (PR babysitter, CI repair, incident response, dependency triage). |

## The one rule that runs through all of them

Whatever you copy, keep these guardrails — they appear in nearly every
battle-tested prompt here:

1. **Fix only the cause; don't widen scope.** Drive-by refactors break loops.
2. **Never weaken or delete a test to make it pass.** The check must stay honest.
3. **The maker doesn't declare itself done — a separate checker does.**
4. **Write progress to a file each turn**, and read it at the start of the next.
5. **Have an explicit stop**: goal met, budget spent, same failure twice, or
   escalate to a human.

See **[docs/05 — Stop Conditions & Verification](../docs/05-stop-conditions-and-verification.md)**
for why these matter.

## A note on attribution & licensing

These prompts were written by their respective authors (linked in each entry).
They're collected here for **education**, lightly normalized, with credit. If
you're an author and want an entry changed or removed, open an issue. This
repo's own text and code are MIT-licensed; the linked sources keep their own
licenses.