Archive / How to hand a repair to an AI engineer without babysitting it

How to hand a repair to an AI engineer without babysitting it

A handoff brief that gives one coding agent ownership of a fix, keeps it working on a copy and brings back a tested change for one review.

For: Solo founders

Topics: Agent management

Why I wrote it

One of my agents broke. Its main conversation had grown so large that every request was refused: the history was several times bigger than the model could read at once. Several AIs looked at it, and I ended up carrying their findings back and forth between them.

I wanted one engineer agent to own the whole fix, so I could stop relaying messages.

What makes it work

It hands over what earlier agents found as "state to pick up", and says to confirm each item before trusting it. Some of those findings had already been wrong once.

All the work happens on a copy. Nothing touches the live system until I've reviewed one proposal.

A small checkpoint file records what's done, the current failure and the next step, so the work survives an interruption.

Routine errors are the agent's job, not new permission requests. It contacts me only for the review, a missing credential, a budget stop or a real blocker.

It defines success up front, including a rollback that has been tested.

The prompt

Prompt

PURPOSE Take ownership of [the repair] on [the machine or system]. Remove the need for me to relay analysis between agents. Produce a tested change ready for one production review. This brief is context, not evidence that any repair has been applied. GOAL Prepare and demonstrate [the fix] on an isolated copy. Preserve history, test rollback, and return the patch and an execution log for review. Work continuously within this scope. HOW TO WORK One engineering owner, one isolated workspace, no other writers. Check what is already running before you start. Keep a small checkpoint file: work done, exact file paths, current failure, next action. Resume from it after any interruption. Use a visible finite budget; no unlimited retry loops. Routine path, dependency and test failures are your work to solve, not new permission requests. Contact me only for the production review, a missing credential, a budget stop, or a blocker you can reproduce after real investigation. AUTHORIZED Read code and config with secrets redacted, make backups, build and test on copies. NOT AUTHORIZED Production deployment, schedule changes, retiring bots, deleting history, sending anything, spending money, changing credentials. STATE TO PICK UP (confirm before using) [What earlier agents reported, numbered. Mark the items that turned out wrong before.] SUCCESS EVIDENCE [What must be shown, for example: full processing, history kept, no duplicates on re-run, correct routing, rollback tested.] Then return one live-change proposal: exact change, backup, test evidence, rollback, remaining limits. Stop there for my review.

How to use it

Give this to a coding agent that can work on your machine, such as Claude Code. Fill in the state to pick up with everything earlier agents told you, numbered, and be honest about which items turned out wrong.

Spend the most time on the success evidence. If you can't say what "fixed" looks like, the agent can't either.

When the proposal comes back, check the rollback before anything else.

Use it for: Fixing a broken agent when several AIs have already weighed in and you're tired of relaying their findings.

Back to the archive