Review Every Diff

Never commit agent output blindly. Even when the agent has been reliable, even when you are in a hurry, even when the change looks trivial — review the diff.

Agents optimize for plausible-looking output, not correctness. They will:

• Rename variables to synonyms that subtly break logic
• Remove comments you wanted to keep
• Add dependencies you did not ask for
• Change formatting in files outside the scope
• Introduce off-by-one errors, race conditions, or security issues

Your review is the safety net. Skip it once and you teach yourself bad habits.

1. Scope creep. Did the agent touch files it should not have?
2. Logic changes. Is the behavior actually what you asked for?
3. Dependencies. Did it add packages, imports, or services?
4. Tests. Did it add, update, or remove tests? Are they meaningful?
5. Security. Are secrets, permissions, or auth handled correctly?
6. Style. Does the output match your team's conventions?

• Use small diffs. Small tasks produce small diffs that are quick to review.
• Set up pre-commit hooks for formatting and linting.
• Run tests before reviewing so you review against a known baseline.
• Ask the agent to summarize its own diff in the response.

Some teams say, "I trust this agent for formatting only." But formatting-only prompts still produce diffs. Still review them. One bad formatter run can destroy a file you care about.

Review the diff before you run tests, not after. Tests can pass on subtly wrong code. Your judgment is the final check.