AI Guides › Workbench
By Nigel Guy · 7 min read
Most people's Claude prompts, and most CLAUDE.md files, keep growing. Every time the model misbehaved you added a line, often in capitals: "CRITICAL", "ALWAYS double-check", "If in doubt, use the tool". On the current models a lot of those lines now work against you. Anthropic's own guide says so, and it names the lines that cause the trouble.
The rule: on the Claude 5 generation, delete the instructions you wrote to make older models try harder, keep the ones that say what you want and why, and test the trimmed version before you trust it.
Anthropic's Prompting best practices page now covers Opus 5 and 5.5, Sonnet 5 and 5.5, Fable 5 and 5.1, and Mythos 5 and 5.1, alongside some 4-series models. Each newer model also has its own short page. One caveat first: I couldn't confirm what Anthropic removed from earlier versions of the guide itself. What I could check is what the current pages tell you to remove from your prompts:
| Cut this | Why, per the docs |
|---|---|
| Shouting ("CRITICAL: You MUST use this tool…") | Newer models respond more to the system prompt, so aggressive wording causes overtriggering. Plain "Use this tool when…" is enough |
| "If in doubt, use [tool]" and other anti-laziness lines | Tools that used to be under-used now trigger properly. These lines make them fire too often |
| "Double-check your answer", "add a final verification step" | Opus 5 already verifies and corrects itself. The Opus 5 page says removing these saves tokens "with no loss in quality" |
| "Show your reasoning" / "write out your thinking" | On Opus 5 and Fable 5 this can be refused under the reasoning_extraction category. Ask for a short explanation instead |
| "Do not think" rules (when thinking is disabled) | On Opus 5 these increase leaked internal tags in the output |
| Long anti-markdown blocks | Fable 5.1 already formats less, so these can strip out structure the content needs |
| Prefilled assistant replies | No longer supported from the 4.6 models on. They return a 400 error |
The Fable 5 page puts it most bluntly: skills written for earlier models are "often too prescriptive" and "can degrade output quality".
| Tool | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Best-practices page | Techniques for all current models, plus migration notes | Free | Anyone writing prompts | Long, and it mixes models. Read the table at the top first |
| Per-model prompting page | What changes on one model, with sample instructions | Free | Moving a prompt to a new model | Advice differs by model. Opus 5 runs long, Fable 5.1 sends fewer progress updates |
/doctor prompt-audit in Claude Code |
Checks your instruction files for lines written for older models, dead references and contradictions | Included with Claude Code on paid plans. Pro is listed at $20 a month in the US; checkout shows your price in pounds | CLAUDE.md, rules, skills, subagents | Needs Claude Code v2.1.283 or later. Proposes edits; it doesn't apply them until you ask |
| The trim-audit prompt (below) | Does the same job for any prompt in ordinary chat | £0 on Free | Custom instructions, Project prompts, API system prompts | Claude's judgement, not a test. You still have to compare outputs |
Find your model in the table at the top, read its own page, then skim Migration considerations. Copy only the sample instructions you need.
/doctor prompt-auditIn a Claude Code session, run:
/doctor prompt-audit
By default it covers CLAUDE.md, CLAUDE.local.md and AGENTS.md, plus the rules, skills, commands, subagents and output styles under .claude/ and ~/.claude/. To audit one area only, pass a path, for example /doctor prompt-audit .claude/skills/deploy. You get a report with proposed edits, and nothing changes until you ask for the edits to be applied. It runs through the bundled /claude-api skill, so it won't work if you've switched bundled skills off.
Paste this into a new chat. Fill in the model you use, what the prompt is for, and the prompt itself.
You are reviewing a prompt I use with Claude. The goal is a shorter version that
behaves at least as well on current Claude models.
Model I use: [MODEL_NAME, e.g. Claude Opus 5.5]
What the prompt is for: [ONE_SENTENCE_PURPOSE]
Problems I have seen recently, if any: [KNOWN_ISSUES or "none"]
<prompt>
[PASTE_YOUR_PROMPT_OR_CLAUDE_MD]
</prompt>
Work through these steps in order:
1. Number every instruction in the prompt.
2. Label each one: KEEP (says what I want or gives context), SOFTEN (right idea,
but shouted or absolute), CUT (pushes for extra effort, re-checking,
tool use "just in case", or asks you to write out your reasoning), or
MERGE (overlaps or contradicts another line - name which).
3. For each SOFTEN, write the calmer version and add a one-clause reason
where the original had none.
4. Produce the trimmed prompt in full.
Output format: a table (number, original line shortened to 12 words, label,
reason), then the trimmed prompt in one code block, then at most three lines
I should test first.
Rules:
- Do not remove safety boundaries, such as asking before deleting, sending or
publishing anything. Mark those KEEP.
- Do not add new instructions I did not ask for.
- If the purpose or model is missing, or a line's intent is unclear, ask me
instead of guessing.
Before you answer, check: is every CUT line one a current model would follow
without being told? Does the trimmed prompt still do the one job I described?
Claude Code's memory docs give you a concrete target:
@path imports help organise a file but don't reduce its cost, because imported files load at launch too.npm test before committing" rather than "Test your changes", and "Use 2-space indentation" rather than "Format code properly"..claude/rules/ with paths frontmatter, so they load only when matching files are open. Put task-specific procedures in skills, which load only when they're relevant./doctor prompt-audit.budget_tokens returns an error on Claude 4.7 and later models. Use the effort setting instead.