Agent Preflight

Run this before letting Claude Code, Codex, Cursor, or another coding agent modify a real repository. The goal is to make the agent produce a short, reviewable safety receipt before any file-changing work starts.

Instructions

1. Identify the proposed agent task

   • Restate the requested change in one sentence.
   • List the files, directories, commands, secrets, services, and package managers the agent is likely to touch.
   • Flag any task that asks for credentials, payment details, KYC, production deploys, destructive commands, private prompt extraction, or unclear external network access.

2. Map the repository blast radius

   • Inspect the project layout, dependency manifests, test commands, CI config, lockfiles, and deployment files.
   • Identify protected paths such as .env*, secrets, auth code, billing code, migration files, infra config, release scripts, and generated artifacts.
   • Separate safe read-only exploration from write operations.

3. Score the change before editing

   • Mark the task Green when it is local, reversible, tested, and limited to low-risk files.
   • Mark it Yellow when it touches dependencies, config, generated code, authentication-adjacent logic, or unclear test coverage.
   • Mark it Red when it may expose secrets, change payment/auth/deployment behavior, delete data, or run untrusted code.

4. Create guardrails for the agent run

   • State the allowed files and commands.
   • State the blocked files and commands.
   • Require a branch or patch workflow for Yellow/Red work.
   • Require tests, lint, or a dry-run command before reporting success.
   • Require the agent to stop and ask before crossing any listed boundary.

5. Produce a preflight receipt

   • Output a concise Markdown receipt with: task summary, risk level, allowed scope, blocked scope, test plan, rollback plan, and exact stop conditions.
   • Do not perform the actual implementation in the same response unless the user explicitly approves the receipt.

Receipt template

# Agent preflight receipt

Task: <one-sentence task>
Risk: Green | Yellow | Red

Allowed scope:
- <files/directories/commands the agent may use>

Blocked scope:
- <files/directories/commands/secrets/payment/deploy actions the agent must not touch>

Required verification:
- <tests/lint/build/manual checks>

Rollback plan:
- <branch, patch, revert, backup, or restore step>

Stop conditions:
- <conditions requiring human approval before continuing>

Optional reference

For a copy-paste scanner pattern and example AI-agent repo-risk receipts, see the free repo: https://github.com/el-zachariah/ai-agent-safety-starter-pack.