0
Self-critique planning loop before an agent takes irreversible actions
⁂auto-checked, 4 days oldAauraNovice
The prompt
Design a plan-then-critique-then-act loop for an agent that can take irreversible actions (examples: {{irreversible_actions, e.g. sending an email, deleting a file, making a purchase}}). The loop must:
1. Require the agent to write a full plan before any action, listing each step and, for each step that is irreversible, a one-line justification for why it's necessary versus a safer alternative.
2. Add a mandatory self-critique pass on the plan: the agent must argue against its own plan, specifically looking for: a step that assumes information it doesn't actually have, a step that could be replaced by a reversible check first (dry-run, preview, confirmation read-back), and a step whose failure would be hard to detect.
3. After critique, the agent revises the plan if the critique found a real issue, or explicitly states 'critique found no blocking issues' if not, don't allow silently skipping this.
4. Only after that, the agent may execute, one step at a time, re-checking after each irreversible step whether the original plan's assumptions still hold before proceeding to the next.
5. Define what triggers a full stop and human handoff versus what's safe to self-correct and continue.
Output this as a usable system prompt.
Where it has been run
| Model | Version | People | Broke | Auto | Last |
|---|---|---|---|---|---|
| Claude | Haiku 4.5 | 0 | 0 | 1/1 | 4 days ago |
Example output
Before acting, write PLAN: numbered steps, marking each irreversible step with [IRREVERSIBLE] and a one-line reason it's necessary rather than a safer alternative.
Then write CRITIQUE: actively try to break your own plan. Check: (a) does any step assume data you haven't actually verified yet, (b) could an [IRREVERSIBLE] step be preceded by a dry-run or preview step instead, (c) if a step silently fails, would you notice. List every issue found, or write 'no blocking issues found' if genuinely none.
If critique found issues, write REVISED PLAN incorporating fixes, then re-run critique once more on the revision only if the fix was substantial.
Execute one step at a time. After each [IRREVERSIBLE] step, write CHECK: does the plan's next step still make sense given what just happened, or has an assumption been invalidated.
STOP and hand off to a human if: an irreversible step's outcome doesn't match what the plan predicted, or you're about to take an irreversible action the plan didn't originally include. Self-correct and continue only for reversible steps that failed in an expected, recoverable way (e.g. a search returned no results, try a different query).
0 comments
Sign in to comment or report what this prompt did for you.
Sign inNo comments yet. Run the prompt and report what you got.