Type /doctor prompt-audit into a Claude Code session and it scans your CLAUDE.md files, skills, agents and commands for prompting patterns written for older models — that is the claude code prompt audit command, and it landed in version 2.1.283 of the official Claude Code changelog, current as of 26 September 2026. The same tool also answers to /checkup prompt-audit, so whichever alias you remember, you get the same report. If you have been building on Claude Code for more than a few months, this is the single most useful housekeeping command Anthropic has shipped this quarter, because the instruction files that power your whole setup quietly rot while you are busy shipping.
📺 Watch: NEW Claude Code Update is WILD!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
Here is the problem it solves. Every serious Claude Code setup accumulates a stack of instruction files: a CLAUDE.md for each project, a library of custom commands you picked up while learning Claude Code, skills for repeatable jobs, and agent definitions for the specialised work. Most of those were written for whatever model was current at the time. Models move on; your files do not. The prompt audit exists to close that gap without you re-reading every file by hand.
What the Claude Code prompt audit command does
According to the Claude Code changelog entry for version 2.1.283, the command was "added to audit your CLAUDE.md files, skills, agents and commands for prompting patterns written for older models". A second entry in the same version sharpens what the report prioritises: "stale paths, stale commands and contradicting instruction files now lead the report, and thinking keywords that Claude Code documents are kept". Read that carefully, because it tells you exactly what the audit hunts for:
- Stale paths — instruction files pointing at directories or files that no longer exist in your project.
- Stale commands — references to scripts, tools or slash commands that have been renamed or removed.
- Contradicting instruction files — two files giving Claude opposite orders, which forces the model to guess which one you meant.
- Old-model prompting patterns — phrasing written for earlier Claude generations that current models no longer need, minus the documented thinking keywords, which the audit deliberately leaves alone.
That last point matters. The audit is not a blunt instrument that strips everything unusual out of your prompts. Keywords that Claude Code itself documents as meaningful are kept, so a tuned setup does not get flattened back to defaults.
| Command | What it checks |
|---|---|
| /doctor prompt-audit | CLAUDE.md files, skills, agents and commands |
| /checkup prompt-audit | Identical — an alias for the same audit |
How to run the Claude Code prompt audit command
The workflow is short, which is exactly why you should actually do it rather than file it under "later":
- Open Claude Code in the project you want to check — the audit reads the configuration that project actually loads.
- Run /doctor prompt-audit (or /checkup prompt-audit — both are listed in the changelog as the same feature).
- Read the top of the report first. Stale paths, stale commands and contradictions lead the output, and those are the findings most likely to be silently damaging your sessions today.
- Fix the flagged files, then re-run the audit until the report comes back clean.
If you run Claude Code across several projects, repeat the pass per project. A shared CLAUDE.md that works in one repository can be quietly contradicted by a project-level file in another, and the contradiction is precisely the kind of thing this audit was built to surface.
If you want your whole Claude Code setup — CLAUDE.md files, skills and agents — built right the first time, the AI Profit Boardroom includes the full Agent OS zip, prompt libraries and weekly live coaching calls → get the working files inside the Boardroom. Prefer 1-on-1 help on your own funnel first? Book a free SEO strategy session and map it out with Julian.
Why prompting patterns written for older models hurt you
The changelog's phrase "prompting patterns written for older models" is doing a lot of work, so it is worth unpacking why this class of prompt debt is worth auditing at all. Older instruction files tend to over-instruct: they repeat the same rule three ways, shout in capitals, and spell out behaviour that newer models already do by default. Every one of those lines still costs context space on every single session, and contradictory or outdated lines are worse than useless — they actively pull the model away from what you want now.
Prompt debt also compounds in the same way technical debt does. Each individual stale line looks harmless, which is why nobody deletes it. But a CLAUDE.md that has survived several model generations can easily be carrying instructions that reference commands you removed months ago. An automated audit is the honest fix, because the one thing you will never reliably do by hand is re-read config files that already "work".
This is also why the audit slots neatly beside the funnel's existing guidance on Claude Code's AGENTS.md support. The changelog shows AGENTS.md support arriving in version 2.1.281 and expanding in 2.1.282 to Amazon Bedrock, Google Vertex AI, Microsoft Foundry and LLM gateways — so the surface area of instruction files Claude Code reads keeps growing. More instruction files means more places for stale instructions to hide, and a bigger payoff from auditing them on a schedule.
📺 Watch: Anthropic Just Put Claude Code in the Cloud
What else shipped in Claude Code 2.1.283
The prompt audit is the headline for everyday users, but the same changelog version carries a set of changes worth knowing about if you run Claude Code in anger:
- An availableModelsMatch managed setting: with the exact option, "an availableModels entry allows only the model version it names, so new releases stay blocked until listed".
- A deniedModels managed setting "to block specific models, even when availableModels allows them".
- MCP tool, WebFetch and WebSearch outputs added to OpenTelemetry span events when tool-content logging is switched on.
- A fix so /context now counts MCP server instructions — "they now appear as their own row and count toward the total", which finally makes context budgeting honest for MCP-heavy setups.
- A fix for stdio MCP servers "being left running when the session ended while they were still starting".
None of those change your day-to-day prompting, but the /context fix in particular pairs well with a prompt audit: first you see what your instruction stack really costs, then you trim what the audit flags.
Where the prompt audit fits an AI SEO and automation workflow
If you use Claude Code to run content and SEO systems — the way the Claude Code SEO agent build does — your instruction files are your production code. A stale path in a hobby project costs you a shrug; a stale command in an agent that publishes for you costs you output quality every single day until you notice. Running the prompt audit before you scale an automation is cheap insurance, and re-running it after any big refactor is cheaper still.
The same logic applies if you are running Claude Code on a budget or off the beaten path: a free-tier Claude Code setup has less context room to waste on rotten instructions, and a local Claude Code configuration tends to accumulate machine-specific paths that break the moment you move house. The audit catches both classes of problem. And if you drive Claude Code through Hermes, the notes on using Hermes Agent with a Claude Code subscription explain how the two stacks share configuration — which is exactly where contradicting instruction files love to breed.
For the bigger picture on structuring agent instructions so they do not rot in the first place, the Agent OS guide covers the file layout this whole funnel runs on, and the Goldie Bench write-up covers how the different model brains compare in hands-on tests — useful context when you are deciding whether an old prompt pattern is still earning its keep on a newer model.
Claude Code prompt audit command FAQ
Is the prompt audit a separate tool I need to install?
No. It ships inside Claude Code itself from version 2.1.283 — you run it as a slash command in an ordinary session, with /doctor prompt-audit and /checkup prompt-audit both working.
Will it rewrite my CLAUDE.md files automatically?
The changelog describes it as an audit of your files that produces a report with the highest-impact findings first. Treat the report as a to-do list you review — you stay in control of what actually changes.
How often should you run it?
A sensible rhythm is after every model upgrade, after any big project refactor, and on a monthly schedule for automation projects that run unattended. The command is fast to invoke, and the report leads with what matters, so there is no real cost to running it often.
If you want Claude Code doing real revenue work — SEO content, automations and agents that run while you sleep — check out the AI Profit Boardroom → join the builders inside. And if you would rather talk through your specific setup first, book a free SEO strategy session — bring your audit report and leave with a plan.
