Brain Audit Checklist

Are all 'Always On' rules still necessary?

Every alwaysApply: true rule consumes context tokens on every conversation. Remove or downgrade to glob/model-decision if not needed globally.

Do glob patterns still match your file structure?

If you've renamed directories or changed file extensions, your glob-activated rules may no longer trigger.

Are any rules being consistently ignored by the agent?

If the agent repeatedly violates a rule, the rule may be too vague. Rewrite it with concrete examples.

Are there rules that overlap or contradict each other?

Conflicting rules confuse the agent. Merge duplicates and resolve contradictions.

Is each rule under the 12,000 character limit?

Antigravity enforces a 12,000 character limit per rule file. Split large rules into focused sub-rules.

Does every SKILL.md have a clear description in frontmatter?

Antigravity uses the description field for skill discovery. Without it, the skill won't be found.

Are skill steps still accurate for your current project?

If your deployment process or testing approach has changed, update the corresponding skill.

Are there manual processes you do repeatedly that should be skills?

This is the Manual Labor trigger from the meta-learning rule. If you do it regularly, codify it.

Are any skills too broad? Should they be split?

A skill that covers 'everything about deployment' is less useful than focused skills for 'deploy to staging' and 'deploy to production'.

Are workflows reflecting how you actually work?

Workflows should be auto-generated from successful completions, not theoretical. If a workflow doesn't match reality, update it.

Are there successful multi-step tasks that should become workflows?

After a complex task succeeds, the agent should propose capturing the steps as a workflow.

Are workflow /commands still discoverable and well-named?

Workflow names should be short, descriptive, and easy to remember (e.g., /deploy, /feature-launch).

Delete any rules, skills, or workflows not used in the past 2 weeks.

If you haven't needed it, you probably don't need it. You can always recreate it.

Merge any near-duplicate rules into a single, focused rule.

Two rules that say similar things waste context. Combine them.

Check total token cost: how many 'Always On' rules do you have?

More than 3-5 Always On rules can start to impact performance. Use glob or model-decision activation for less critical rules.

Has the agent proposed any upgrades this week?

If not, either your system is very mature or the meta-learning rule isn't triggering. Test it with a deliberate repetition.

Were proposed upgrades actually useful?

If the agent is proposing low-quality upgrades, the meta-learning rule may need tightening.

Are you approving or rejecting upgrades with clear reasoning?

The human-in-the-loop principle means every upgrade needs explicit approval. Don't auto-accept.