Anthropic has added the /claude-api prompt-audit command to the official claude-api skill for Claude Code. It looks for “anti-patterns” from the era of older models in CLAUDE.md, skills, agent.md, tool descriptions, and prompts, and suggests removing them. According to the company’s benchmark on a customer support task, migrating from Opus 4.8 to Opus 5.5 alone saved about 18% of tokens, while prompt cleanup added another ~9% savings and about 2 percentage points of accuracy. Auditing harness configurations is shifting from optional hygiene to a standard step in moving to a new generation of models.

!image.jpg)

image

What happened

The /claude-api prompt-audit command appeared in the official claude-api skill for Claude Code. It scans CLAUDE.md, skills, agent.md files, tool descriptions, and prompts, flags typical anti-patterns from the era of previous models, and suggests removing what it finds. The list includes verification rituals like “double-check your work,” all-caps emphasis in the style of “CRITICAL: YOU MUST,” forced scratchpad reasoning, conflicting rules, and outdated thinking settings. The company shows the quantitative effect in a blog post using a benchmark for the Opus 4.8 → Opus 5.5 migration on a customer support task. The same post breaks down specific failure mechanisms: an outdated thinking setting caused API requests to fail, and forced manual scratchpad reasoning led the model to write a tool call inside its reasoning without executing it.

Context

Verification rituals, all-caps emphasis, and forced intermediate reasoning previously solved a real problem: they compensated for weaknesses in older models that missed errors or lost the thread in multi-step scenarios. For frontier models with built-in reasoning, these crutches are no longer neutral — they consume tokens, break prompt caching, and conflict with the model’s own reasoning mechanism, while the model executes legacy instructions literally and to its detriment. This leads to a less obvious conclusion: harness configuration — prompts, skills, CLAUDE.md — is effectively a dependency on a specific model snapshot, not neutral text that can be left untouched for years. The context for the command’s release is the launch of Claude Opus 5.5 on September 22, 2026, which, according to official release notes, is 40% cheaper to operate than Opus 5: as token prices fall, the share of spending on outdated instructions becomes more noticeable, and the benefit of cleaning them up becomes more measurable.

Why this matters for the industry

The key industry signal is the proportion of the effect: according to Anthropic’s benchmark, cleaning up configurations yields a benefit comparable to half the gain from a model upgrade. In other words, accumulated “config debt” in prompts, skills, and CLAUDE.md is a measurable line item in costs, not abstract hygiene. Harness auditing is being built into migration playbooks alongside eval runs, and Anthropic’s decision to embed the check inside an ecosystem skill makes it the default for the Claude Code ecosystem. A likely development is the formation of a separate tooling layer: prompt linters for third-party harnesses, configuration checks in CI, and versioning of prompt configs with regression tests against specific model generations. Meanwhile, teams have to recalculate the unit economics of agentic features: the Opus 5.5 discount combined with prompt cleanup creates a margin cushion that did not exist before.

Why this matters for users

Anyone working with Claude Code or another harness can test the effect themselves: the command can be run on their own repositories right now, before and after upgrading to Opus 5.5. The token savings and accuracy gains claimed by Anthropic should be validated on their own tasks — measure tokens per task, the share of tool calls actually executed, and answer quality before and after cleanup, rather than relying only on numbers from someone else’s case. A separate reason for manual review is thinking settings and forced scratchpad instructions: according to the company’s examples, these most often cause API and tool failures, and such breakages are easy to miss if you only monitor final responses.

What is still unknown / limitations

The evidence base is still thin: the numbers come from a single vendor case — a customer support task during the Opus 4.8 → Opus 5.5 migration. The published materials do not include sample size, variance, or definitions of the “savings” and “accuracy” metrics, so the result should correctly be treated as a hypothesis to be tested with your own evals, not a guaranteed effect for any tasks, models, or harnesses. The conclusion about the universal necessity of auditing is currently no better supported than a single case: reproducibility of the numbers on other tasks, with other vendors, and in open-source harnesses remains to be confirmed by independent measurements.

Sources

Author

Look at AI, editorial team