Free template
AI Agent Setup Runbook
Use this one-page runbook before letting an agent work. It forces four decisions: what the task is, what the agent may touch, how you verify the result, and when it must stop.
One-page template
Copy this into your repository. Replace the examples with one real task.
# Task runbook
## Task
- Goal: Revise one module / create one report / summarize one source.
- Input files: `src/example.py`, `docs/brief.md`
- Desired output: `docs/output.md`
- Verification: `npm test` / `python script.py --check`
## Safety boundaries
- May read: files listed above.
- May write: only the desired output.
- May execute: only the verification command.
- Scope check: run the declared write scope through an offline preflight (e.g. Agent Role Contracts) before dispatch; if the runtime allow-list is broader than the declared scope, classify the gap as an integration issue, not a tool defect.
- Must ask first: deleting files, installing dependencies, sending messages.
- Never: credentials, customer data, production systems.
## Stop rule
- Stop after two failed attempts.
- Stop if the task expands beyond the listed files.
- Stop if verification output cannot be shown.
## Review record
- Reviewer:
- Result:
- Evidence:
- Next step:Why this works
Less drift
The agent has a fixed input, output, and verification command.
Clear limits
Permissions are explicit instead of inferred from a long chat.
Reusable
The same form works for code, research, reporting, and document work.
Need it configured for your repository?
In a 45-minute session, I adapt this runbook to one real task, run one representative test, and leave you with the working version.