A lightweight preflight checklist before trusting an AI coding agent

Most AI-coding failures I see are not model failures. They are missing preflight checks: unclear repo state, unreviewed permissions, no rollback path, and no tiny verification step after the agent finishes.

Here is the lightweight checklist I now use before I let an agent touch a project:

1. Define the one allowed change

Write one sentence: "The agent may only change ___ to achieve ___." If you cannot write that sentence, the task is probably still too vague.

1. Freeze the starting state

Before the agent starts, record:

• current branch and commit
• whether the working tree is clean
• which files or directories are in scope
• which production systems are off limits

This avoids the classic "it also cleaned up a few unrelated things" problem.

1. State forbidden actions explicitly

Examples:

• do not deploy
• do not touch billing or payment config
• do not read private customer data
• do not rotate keys
• do not rewrite history
• do not change prompts or security policy

For agents, an unstated boundary is often not a boundary.

1. Require a rollback note before execution

The rollback does not have to be fancy. It can be:

• revert commit X
• restore file Y from git
• disable feature flag Z
• re-run previous deployment command

If there is no rollback path, the task is probably too large for one agent turn.

1. Use the smallest meaningful verification gate

Pick one real check:

• unit test
• typecheck
• build
• lint
• screenshot
• direct browser check
• curl against a health endpoint
• manual inspection of the exact changed file

Do not accept "looks good" if no gate was run.

1. Separate read-only research from write actions

An agent can search docs and inspect logs in a broad way, but writes should be narrow and auditable. I treat external posting, account settings, production deploys, billing, privacy data, and scheduler changes as separate permission zones.

1. Keep a short handoff

At the end, ask for:

• what changed
• what was not changed
• what check passed
• what remains risky
• exact files touched

I am experimenting with a small readiness workflow around this for early AI-agent projects: a preflight card that turns a vague task into scope, forbidden actions, rollback, and verification before the agent starts. Not a full platform pitch yet; I am mainly validating whether this checklist catches the mistakes teams actually care about.

If you use coding agents in a real repo, what is the one preflight check you wish you had enforced earlier?