Clean my AI harness.

The setup you built without noticing

A harness is everything wrapped around the model: your custom instructions, project files, saved prompts, memory, skills, tools, permissions, and checks. Every time the AI missed something, you added a rule. Each rule fixed a real problem.

Over time the model ends up working inside a system nobody can see in one place. That matters most when the model changes, because the new model inherits the old setup. When the experience gets strange, most people add another rule, and the harness grows again.

This skill breaks that loop. It makes the accessible setup visible, separates the model from the machinery around it, and gives every useful control a clear job.

Claude edition

Download the Claude edition ↓

In Claude, turn on Code execution and file creation under Settings > Capabilities. Open Customize > Skills, click +, choose Create skill, then Upload a skill. Upload `clean-my-ai-harness-claude.zip` and enable it. Then start a new chat on the project you want to review and paste this:

<prompt>
  <task>
    Use clean-my-ai-harness to review the AI setup for this project. Start read-only. Show me one plain-English report. Do not change anything until I approve each change.
  </task>
</prompt>

If your instructions live outside the project, add or export them before running. To remove the cleaner, disable or delete it under Customize > Skills. That does not undo changes you already approved; use that run's rollback.

Codex edition

Download the Codex edition ↓

Requires Python 3.10 or newer. Unzip `clean-my-ai-harness-codex.zip` and copy the `clean-my-ai-harness-codex` folder into `~/.codex/skills/` (create it if it doesn't exist). Restart Codex. Then, in the repo you want to review, invoke it by name:

<prompt>
  <task>
    $clean-my-ai-harness-codex review the AI setup for this repository. Run your read-only scanner and generate the report. Read-only means do not modify my existing files, but you should run the scan and create the report. Do not apply any change until I approve it.
  </task>
</prompt>

To remove the cleaner, delete `~/.codex/skills/clean-my-ai-harness-codex`. That does not undo changes you already approved; use that run's rollback.

Say the one sentence

Open the project you want to review and paste the sentence below. The first run is read-only. The cleaner treats everything it inspects as untrusted data: it never follows instructions found inside your files, runs your scripts, opens your links, or widens its own scope because a file tells it to. Nothing is edited, moved, or installed until you approve it.

<prompt>
  <task>
    Review the AI setup for this project. Start read-only and show me what you would keep or change.
  </task>
</prompt>

Open YOUR-AI-SETUP.html

The cleaner returns one visible report, `YOUR-AI-SETUP.html`. It shows what shapes your AI before you type, what protects the work, what creates confusion or extra work, what the cleaner recommends, and what it could not see.

The detailed evidence stays in a hidden folder the cleaner manages for you. You never touch file lists, hashes, or schemas.

Blind spots are named, not hidden. If the cleaner could not see something, the report says so instead of guessing.

Six things the cleaner can propose

Each control it reviews gets exactly one recommendation: Keep it, Give it one home, Load it later, Turn it into a check, Put it on probation, or Retire it safely.

For each one you see what you may notice now, what would change, the evidence, what must survive, the risk if it's wrong, and how to roll it back. Long is not the same as bad, and repeated wording is not automatically redundant.

When the evidence can't support a clean decision, the cleaner records a coverage gap instead of guessing.

What the Claude edition looks at

How much procedure arrives before Claude has seen the job, when specialist material joins the work, which action boundaries must stay explicit, and whether exact delivery requirements have a real final check. It reads the Fable 5 model profile only when Fable 5 is in scope.

What the Codex edition looks at

Discovery and routing first: whether Codex can find the right skill, whether several skills claim the same job, whether the full procedure stays behind selection, and whether schemas, permissions, tests, and validators carry the yes-or-no requirements. It reads the GPT-5.6 model profile only when that model is running.

Approve the ones you want

When you're ready, reply with the numbers. A vague reply never approves the batch, and approving one item never approves the rest. Before it applies anything, the cleaner re-checks that the reviewed setup still matches what it scanned, then applies only the items you marked. If a number is unclear, it asks about that number only.

<prompt>
  <task>
    Approve 1 and 3. Leave 2.
  </task>
</prompt>

What changed, and how to undo it

After an approved cleanup you get one visible `WHAT-CHANGED.md`: what moved, what stayed protected, what passed, and how to undo it. The run receipt records the checks and a final status: `REVIEW REQUIRED`, `APPLIED AND VERIFIED`, `PARTIAL FAILURE`, or `NO SAFE CHANGE`. Rollback restores the exact starting state.

What “better” means

A cleanup is better when the important rules still work, repeated ownership is reduced, specialist material no longer arrives too early, yes-or-no requirements have real checks, rejected and unrelated files stay untouched, and the setup can be restored.

It does not promise a fixed time saving. A structural cleanup is not a measured performance result until you test the same work before and after.

Re-run after a major model release, or when the same new failure keeps showing up. The harness grows again; a regular re-audit keeps it honest.