A practical preflight checklist for AI coding agents

A pattern I keep coming back to with AI coding agents is: the agent is usually not the risky part — the missing preflight is.

Before I let an agent touch a real codebase, I try to answer five questions:

1. What is the smallest reversible change?
2. Which files are off-limits?
3. What command proves the change did not break the product?
4. What evidence should the agent collect before saying “done”?
5. What is the rollback plan if the result looks plausible but wrong?

For example, instead of asking an agent to “add billing,” I would split it into:

• inspect the existing payment flow and summarize assumptions;
• add one narrow test or fixture first;
• implement one path behind a flag;
• run the smallest relevant test/build;
• report exact files changed, commands run, and remaining uncertainty.

The useful habit is to treat agent work like a pull request from a very fast junior teammate: give boundaries, require evidence, and review the diff rather than the confidence.

A small template I use:

Goal:
Non-goals:
Allowed files:
Risky areas:
Validation command:
Rollback:
What to report if blocked:

This adds a few minutes up front, but it saves a lot of time debugging confident, half-correct changes later.

Curious how others are handling this: do you use a formal preflight checklist before agent-assisted coding, or do you keep it lightweight and task-specific?