“AUDIT” The One Command That Fixes Your Outdated Claude Code Skills

Your skills are stale and you did not notice

Every skill you wrote six months ago was tuned for the model you had six months ago. That model needed to be shouted at. It needed "ALWAYS" and "NEVER" and "CRITICAL" in caps. It needed step-by-step scripts because it planned badly on its own.

Fable 5.1 does not need any of that. It follows a plain instruction the first time. So all that pressure language now does the opposite of what you wanted. The model over-triggers, gets rigid, and writes cautious, hedgy output because your prompt sounds anxious.

Worse, skills grow. You add a rule after a bad session. You paste in a second reference file. Six months later the skill has two sets of rules that disagree and the model picks one at random each run. That is why the same skill gives you a great result on Monday and slop on Tuesday.

What the command actually does

Claude Code ships with a skill called /claude-api. One of its subcommands is prompt-audit. It reads your skill file and every reference file it loads, then scans for four kinds of cruft:

Then it does the thing most cleanup tools get wrong. It has a keep list.Your audience, your offer, your quality bar, your business constraints: it leaves all of that alone. The audit is not "make it shorter." It is "find the specific lines that hurt."

It proposes every change as a diff. It applies nothing until you say which ones.

What it found on mine

Opt-in page skill: 11 problems.The big ones were shouting in the rules section and a pile of layout instructions that fought the model's own design sense. After the cleanup the pages went from generic AI layout to something I would actually ship.

Copy skill: 5 problems, 3 of them direct contradictions.One reference file said "ask 6 intake questions." The main file said 4. One said "for mixed audiences default to the dominant gender." The other said "lead female-safe." The model had been flipping a coin on every run and I had no idea.

The token savings were small. Around 5 percent. The win was consistency.Same questions every time, same close every time, no random "want three more angles?" tail on every deliverable.

Paste this into Claude Code

Swap in your skill name where it says [INSERT SKILL].

/claude-api prompt-audit my [INSERT SKILL] skill. After you've done the audit, create a beautifully designed html file to help me visually understand the main things you would change. On the HTML file, make sure your proposed changes are as easy to understand as possible.

Design rules for the HTML file:
- Open with a one-sentence thesis and the single most important number in a solid accent block.
- Show a proportional bar: tokens removed vs tokens that stay.
- One card per change, ranked by output effect (not by token savings). Each card gets: a large number, a plain-English title, the file and line range, an impact pill (High / Medium / Low), a two-pane diff with the removed text struck through on the left and the exact replacement on the right, and a footer with WHY (what pattern it matches and why it hurts on current models) and OUTPUT EFFECT (what I would notice afterward, no hype).
- Inside the before pane, annotate the specific conflict, e.g. "SKILL.md line 74 says the opposite."
- A "Left alone on purpose" block listing what matched a pattern but was kept and why.
- A footer with the recommended order to apply the changes and a note that nothing has been applied yet.
- Dark ground, one accent color, light mode as a real second palette. One display font and one monospace font from Google Fonts. Red tint plus strikethrough for removed text, green tint for replacements, nothing else uses red or green. Cards fade in on scroll, disabled under prefers-reduced-motion. Responsive, no horizontal scroll, no emoji, no gradient hero, no placeholder text.

The first two sentences are the whole audit. Everything after "Design rules" only shapes the report. Do not add rules about what to find. The audit already does that well on its own.

How to read the report

Every card answers three questions.

Before and after.Red is the exact text that goes. Green is the exact text that replaces it. If the after pane says "removed," nothing replaces it.

Why. Which pattern it matched and why that pattern hurts on current models. If the why does not convince you, skip the card. You know your skill better than the audit does.

Output effect.What you would actually notice. This is the honest column. Some changes say "no visible change, pure hygiene." Take those or leave them. The ones that say "fixes a contradiction" are the ones to take.

Then look at the "Left alone on purpose" block. If something you care about is missing from it and also missing from the change cards, the audit did not see it. Ask.

Apply the changes

Say which numbers. "Apply 2, 3 and 5." Claude edits the files and nothing else.

• Do the contradiction fix first. It has the biggest effect and the lowest risk.
• If a change deletes a folder or a whole file, confirm that one on its own. Move it to Trash, never delete outright.

Then run the skill on a real task and compare. If something got worse, put that one line back in its simplest form. Do not paste the old shouting version back in.

When to skip it

• Skills that are mostly exact commands. Deploy scripts, auth flows, anything where one wrong step breaks something. Those are supposed to be rigid. The audit knows this, but read those cards twice.
• Rules you wrote on purpose after a real failure.If the model still makes that mistake today, keep the rule. The audit asks "does this failure still happen," and only you know.
• Skills you wrote this month. Nothing has rotted yet. Run it in six months.

Nothing to install

/claude-api ships inside Claude Code. Type it in any session and it is there.

If it does not show up, update Claude Code to the latest version and reopen your session.

All Resources