Most CLAUDE.md files are full of instructions written for older Claude models: double-check everything, IMPORTANT in capitals, step-by-step scripts for things the model already knows how to do. Opus 5.5 follows instructions more closely and more literally than the models those lines were written for, so the leftovers now cost you. It does the extra checks, over-applies the shouted rules and over-plans, and you pay for every token of it.
Anthropic measured it. On a customer-support benchmark moved from Opus 4.8 to Opus 5.5, removing those old patterns cut costs by a further 9%, and accuracy went up by about 2 percentage points on the same model. Anthropic has also built a command into its own Claude Code skill that finds them in your files and proposes the fix. This guide explains what it looks for and how to run it.
Read Anthropic's post, Reducing cost and improving performance with Claude Platform: https://claude.com/blog/reducing-cost-and-improving-performance-with-claude-platform
Why old instructions make a new model worse
Older, less steerable models needed forceful prompts, so people wrote CRITICAL, MUST and ALWAYS, and scripted every step. Current models respond closely to what you tell them, so the same text over-applies. A rule in capitals fires in places it was never meant for. A 'check your work twice' line buys a second pass of tool calls and reasoning you did not need. A step-by-step script makes the model follow your plan instead of its own, better one. The cost dropped in Anthropic's test because those extra tool calls and duplicated reasoning went away.
The three habits it hunts
| Habit | What it looks like | What to write instead |
|---|---|---|
| Verify-twice rituals | "Double-check your work." "Verify twice before responding." | Delete it. If a check really matters, name the check: "Run the tests before you commit." |
| Orders in capitals | "CRITICAL: YOU MUST ALWAYS...", "IMPORTANT: NEVER...", "Be maximally thorough!!" | Say it once at normal volume, with the reason: "Use this tool when X, because Y." |
| Step-by-step scripts | "Think step by step." "Step 1: read. Step 2: plan. Step 3: ..." | Describe the goal and the quality bar, not the method. Keep exact steps only where one sequence is safe. |
It also flags stale examples written for older models, rules that contradict each other, old model names and dated request settings in code that calls the API.
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (4 more sections, 1 copy-paste prompt) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
Unlocks: Who built it · Run the audit · What it keeps · Before you rely on it
50,000+ follow where these guides come from. No spam, reply “unsubscribe” any time.