New tool cleans up Claude prompts for Opus 5.5

What's new

Anthropic has updated a built-in feature in the developer tool Claude Code. It is called /claude-api prompt-audit and is part of the company's claude-api skill (a collection of helper functions). The command is a subcommand in Anthropic's claude-api skill that scans prompts, skills, CLAUDE.md files, tool descriptions and code, looking for instructions written for older models. The tool was updated when Claude Opus 5.5 came out on September 22, 2026, and now removes what makes the new model waste time and money.

What's clever

Each issue in the report shows the file and line, the quoted text, the pattern, why it is outdated, a severity level (high, medium, low) and the suggested fix. It checks six known traps: verification rituals, thoroughness and emphasis-booster phrases, forced procedures and draft templates, outdated examples and conflicting rules. The tool does not change anything itself; it runs without asking along the way (read-only) and without requiring interaction. It gives the owner a proposal that must be approved manually.

Cheaper or better?

The effect has been measured by Anthropic itself. On a customer service assistant, a test showed, without changing models, a further 9 percent reduction in costs and about two percentage points better accuracy, just by using the audit tool. This comes on top of Opus 5.5 itself being cheaper: the model costs $4 per million input tokens and $20 per million output tokens, 20 percent below Opus 5, and cache reads now cost 60 percent less, down to $0.20 per million tokens. Overall, Anthropic expects typical workloads to become 40 percent cheaper and the model to deliver output more than 30 percent faster than Opus 5. Competitively, Opus 5.5 scores 66.4 percent on the Terminal-Bench 4.0 benchmark at extra-high effort, clearly above OpenAI's GPT-6 Astra at 57.9 percent, and even at standard effort Opus 5.5 beats the competitor's top score while costing about 80 percent less per task.

What it's good at

The audit tool is best suited for companies and developers who have used Claude (or other AI models) for a long time and have accumulated old habits in their prompts and CLAUDE.md files. One user of the tool said the biggest savings were not in the text itself: the most expensive anti-patterns were not in CLAUDE.md but in the tool descriptions, and cutting 12 tools down to 4 fixed more than any prompt rewrite. That makes the feature especially relevant for companies running agent-based customer service, coding or other automated AI tasks at large scale, where even small text changes can yield noticeable savings in hard cash.