← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Give Your AI Instructions
A Checkup

Your model got smarter. Your saved rules didn’t. Here’s how I audit mine in Claude Code and Codex, and how to put it on a schedule so it only ever suggests.

Most of the rules in your setup were written to fix an older model. Anthropic’s own prompting guide says newer Claude models can overtrigger on old pushy tool rules, and its fix is to swap "CRITICAL: You MUST use this tool when..." for a calm "Use this tool when...". OpenAI’s GPT-5 prompting guide says the same about a "Be THOROUGH" rule that helped older models and turned counterproductive.

This page is the setup for both tools plus the schedule. For the full cleanup routine, use my Claude Code maintenance checklist. If you’re still building your first CLAUDE.md, start with the most important file in Claude Code.

Claude Code: Run The Prompt Audit

01

Update first

The audit needs Claude Code v2.1.283 or later. In your terminal, type claude --version, and if it’s lower, run claude update. (Update docs, changelog)

02

Run it

Run: /doctor prompt-audit

Type it in your project folder. /checkup prompt-audit does the same thing. Plain /doctor is a different checkup for your install, settings and plugins. (Commands list)

By default it reads your CLAUDE.md, CLAUDE.local.md and AGENTS.md files, plus the rules, skills, commands, subagents and output styles in your project’s .claude folder and your personal ~/.claude folder. It looks for rules written for older models, files or commands that don’t exist anymore, and files that contradict each other. (Audit docs)

To check one skill, add its path, like /doctor prompt-audit .claude/skills/deploy. The notes Claude saves on its own aren’t on that list, so open /memory and skim those yourself.

03

Approve what changes

You get a report with proposed edits, dead paths and clashing files at the top. Nothing changes until you ask, and I approve by number, never with a blanket yes.

Approve Some Edits
Before you change anything, show me the exact before and after for items [1, 3 and 5]. Then apply only those and leave everything else alone. Tell me which file each change landed in so I can undo it if I need to.

If the command says it’s unavailable, the audit runs through the built-in /claude-api skill, so check you haven’t turned that skill off in your settings.

Codex: Audit AGENTS.md And Skills

01

Know where Codex reads from

  1. Your global file. ~/.codex/AGENTS.md, or AGENTS.override.md in the same folder if it exists. That one wins. (AGENTS.md docs)
  2. Your project files. One file per folder, from the top of your project down to the folder you’re working in. The closer file wins, and Codex stops reading at 32 KiB total, so a bloated file up top can push the closer files out.
  3. Your skills. .agents/skills in your project and ~/.agents/skills for you personally, each skill a folder with a SKILL.md. (Skills docs)

02

Paste the audit prompt

Open a Codex chat in your project, in the ChatGPT desktop app or the Codex CLI, and send this.

Codex Instructions Audit
Audit every instruction you load for me: my global AGENTS.md in ~/.codex, every AGENTS.md and AGENTS.override.md in this project, and every SKILL.md in .agents/skills and ~/.agents/skills. Flag rules written to fix an older model, ALWAYS or all-caps rules a newer model will over-follow, references to files, folders or commands that no longer exist (check before you call one missing), rules that contradict each other, and the same rule repeated in more than one place. For each finding, quote the file and line, say why it's a problem, and propose the exact rewrite or deletion. Number them with the riskiest first. Don't edit anything until I reply with the numbers I approve.

Reply with the numbers you approve, and keep the chat open for the schedule step.

Schedule It

The Safe Rule

A scheduled audit suggests. It never edits. You read the report and approve changes yourself, because a rule that looks stale might be the one thing holding your setup together.

My cadence is the 1st of every month, plus a manual run the day I switch models. Skip /loop for this one. It only runs while your session is open, and repeating loops end after 7 days. (Loop docs) My /goal, /loop and /schedule cheat sheet and autopilot guide cover the rest.

01

Claude Code desktop task

I use this one because it runs on your computer and sees your personal ~/.claude files. In the Claude desktop app’s Code tab, open a session in your project and send this.

Monthly Prompt Audit
Set up a local scheduled task in this folder called monthly-prompt-audit. It should run /doctor prompt-audit at 9am on the 1st of every month, in Plan mode so it can't change any files. Each run should end with a numbered list of findings and proposed edits, riskiest first, and then wait for me. Run it once now so I can see the report.

Then click Scheduled in the sidebar (older apps say Routines) and open the task to check the mode reads Plan. Plan mode explores and proposes without editing your files. (Mode docs) To build it by hand instead, click New task, type /doctor prompt-audit as the instructions, turn on Only on this computer under Advanced settings, pick your folder, then use the pencil icon to set Custom to 0 9 1 * *.

It runs while the app is open and your computer is awake, with one catch-up run if you missed it. Open a run from History to read the report and approve by number. (Desktop task docs)

02

Claude Code cloud task

For instructions that live in a GitHub repository, like a team’s shared CLAUDE.md. It runs with your laptop closed, on a fresh copy of the repository, so your personal files aren’t there. Type this in Claude Code.

Cloud Monthly Audit
/schedule a monthly cloud task for my [REPO NAME] repository that runs /doctor prompt-audit at 9am my time on the 1st of every month. Report the findings and proposed edits only. Don't edit files, push branches or open pull requests.

Manage it at claude.ai/code/routines. If you type a custom schedule there yourself, the hours are in UTC. (Cloud task docs)

03

Codex scheduled task

Back in the chat where you ran the Codex audit, send this.

Codex Monthly Audit
Turn the audit you just ran into a scheduled task for this project. Run it at 9am on the 1st of every month and start a new chat for each run. Report findings and proposed edits only, and never edit, delete or move a file during a scheduled run.

Runs land under Scheduled in the ChatGPT desktop app sidebar, and the app has to be open with your computer on. Runs use your default sandbox setting, and in read-only mode any file change fails, which is the strongest version of the safe rule. (Scheduled task docs)

What To Fix First
  1. Dead paths. Deleted files, renamed folders, paths from an old laptop. The AI wastes a step hunting for them every time.
  2. Pushy tool rules. "ALWAYS use the search tool" was a fix for a forgetful model. Rewrite it as when to use the tool.
  3. Old "double check everything" lines. Newer models check their own work, so these turn into extra rounds you pay for.
  4. Two rules that disagree. Usually global against project. The Claude Code docs warn it may pick one at random.
  5. A CLAUDE.md and an AGENTS.md side by side. By default Claude Code reads only the CLAUDE.md, so rules in AGENTS.md get ignored. Add the line @AGENTS.md to your CLAUDE.md and both tools share one file. (How to share one file)
  6. The same rule in three places. Keep one copy in the file that loads where it matters, and delete the rest.
  7. Anything naming an old model. If a rule says why it exists and that reason is gone, it goes too.

Not sure about one? Hide it for a session and see if anything breaks. That’s the deletion test in my maintenance checklist.