Agent Recovery Checklist Prompt

Start here

Paste this prompt into any ordinary AI chat that accepts text. Below the prompt, paste the five input blocks it asks for: the original task and success condition, the step-by-step tool log, whatever partial output exists, your constraints, and the blast radius (anything already sent, saved, or scheduled).

You do not need a terminal, an API key, or any account connection. The assistant only organizes what you paste; it cannot access your tools or retry anything.

After it replies, check two things before acting: every item under its verified list is something you personally opened or saw, and every retry step has a pass/fail check you can actually run.

What this prompt does

Agent tasks often fail halfway: a tool times out, a file comes back empty, or output looks complete but was never checked. This prompt forces a structured recovery plan instead of a vague suggestion to try again. It sorts the failure into one of a few types, splits results into verified and unverified, flags possible side effects, and gives each retry a precondition and a simple check. It also states when to stop rather than loop.

Before you paste

• Gather the real failure text where you have it, not a paraphrase, since vague input produces vague classification.
• Decide what must not be re-sent or re-run, because duplicate actions are the biggest risk here.
• Note any deadline, since it changes whether recovery or a manual redo is worth it.

Permissions and limitations

• The prompt claims no tool, account, or file access and performs no retries, sends, or scheduling.
• It is a planning aid for general tasks; it does not give legal, medical, or financial advice.
• Anything you have not checked yourself is labeled unverified by design. Do not treat that as a guarantee it is wrong, only as unchecked.
• If the task is small enough to redo safely, the prompt says so rather than building an elaborate plan.
• Source reviewed; runtime not tested.

Privacy

Your tool log and partial output may contain names, customer data, or file contents. Paste only what the assistant needs to plan, and consider removing secrets, credentials, and personal details first.

FAQ

What if I do not have a full step log? Say so in the inputs. The prompt should mark unclear failures and ask for the one or two facts that would settle the classification instead of inventing logs.

A retry step looks safe but might duplicate a sent message. What then? Treat it as risky and confirm manually first. Duplicate sends are the main reason the prompt insists on a side-effect check before the checklist is executed.

Attribution

Original TokRepo prompt, licensed CC BY 4.0. Reference context: ChatGPT release notes, reviewed 2026-10-04; these templates do not depend on a specific announced feature.

Complete reusable prompt

You are helping me recover an Agent task that has partially failed. I will paste the details; you will produce a recovery checklist. You do not have access to my tools, accounts, or files, and you cannot retry anything yourself. You only organize what I give you into a plan I can execute manually.

INPUTS I WILL PASTE

1. Original task: what I asked the agent to do, and the success condition.
2. Tool/step log: each step, the tool used, and its result (success, error, timeout, or partial output).
3. Partial results: any output produced so far, even if incomplete or unverified.
4. Constraints: deadline, permission limits, anything that must not be redone or re-sent.
5. Blast radius: whether anything was already written, sent, scheduled, or changed outside my review.

WHAT TO DO

• Classify the failure: tool error, timeout, bad or empty input, permission denial, output that is partial but plausible, or unclear.
• Separate verified from unverified partial results. Mark anything I have not personally checked as unverified.
• Build a step-by-step recovery checklist: what to inspect, what to retry, what to redo from scratch, and what to discard.
• For each retry step, state the preconditions and a simple pass/fail check I can run before moving on.
• Flag side effects that may already exist (sent message, saved file, scheduled item) and what to verify or undo manually.
• State clear stopping conditions: when to abandon, escalate, or ask a human instead of retrying.

OUTPUT FORMAT Return Markdown with these sections:

1. Failure Classification - one line, plus confidence (high/medium/low) and why.
2. Verified vs Unverified Results - two short lists.
3. Side-Effect Check - items I must confirm or undo before anything else.
4. Recovery Checklist - numbered steps, each with precondition and pass/fail check.
5. Stop and Escalate Rules - when not to retry.
6. Open Questions - facts I must supply before the checklist is complete.

UNCERTAINTY AND MISSING INPUT

• If the failure type is unclear, say so and list the one or two facts that would settle it.
• If partial results cannot be verified from what I gave you, label them unverified. Never assume an unseen output is correct.
• If a retry could duplicate an action, mark it as risky and require my confirmation first.
• Do not fill gaps with invented tool behavior or invented logs. Ask instead.

REVIEW CHECKS BEFORE ACTING

• Confirm every verified item was actually seen by me, not just reported as success.
• Confirm no step re-sends, re-posts, or re-schedules something that may already exist.
• Confirm each retry has a pass/fail check, not just an instruction.
• Confirm stopping conditions are stated for every retry loop.

BOUNDARIES

• You prepare a checklist only. You do not perform retries, send messages, schedule work, or access accounts.
• Do not give legal, medical, or financial advice; if the task touches those areas, tell me to consult a qualified person.
• If the task is trivial enough to redo from scratch safely, say so briefly instead of over-engineering recovery.

WORKED EXAMPLE (fictional, for shape only) Original task: download 12 invoice PDFs and compile a totals table. Tool log: 9 succeeded, 3 returned timeouts, 1 returned a file with 0 bytes. Partial result: a table with 9 rows and one blank. Blast radius: nothing sent or saved externally. Expected output shape: classification = tool timeout plus one empty file, confidence medium; verified = 9 PDFs I opened; unverified = the totals table; side-effect check = none; checklist = re-open the 0-byte file, re-run the 3 timeouts one at a time with a size check over 1 KB, rebuild the table after all 12 exist; stop rule = abandon if a second timeout yields the same empty file; open question = did the source list change.

References and reuse

• ChatGPT release notes · Reviewed 2026-10-04

Original TokRepo prompt · CC BY 4.0. Reference documents retain their own rights.