The model has upgraded, but your configuration files are still stuck in the previous generation.

Lance Martin from Anthropic released a one-line Claude Code command:

```

/claude-api prompt-audit

```

It scans your `CLAUDE.md`, `AGENTS.md`, and `skills` folder to identify prompting habits written for older models — habits that are currently holding Opus 5.5 back.

We recommend running it once before your next session.

There is a passage in the screenshot that tells the story of Great Waves, a sumo wrestler: He never lost a match in private practice, but lost every time he competed in official tournaments. His master told him to stay in the temple and become the great wave his name represents. The next morning, he defeated all his opponents, and never lost a match again.

The meaning of this story here is: Stop being the person "who expects to lose".

### What's the problem?

The context window of Claude Code has long gone from "a major limitation" to "not a problem at all", but countless `CLAUDE.md` files are still full of defensive instructions written for the constraints of the old era.

One developer shared that he ran a similar audit last month, and found half of his "think step by step" instructions were just noise that the model completely ignored. Another person asked: Does the audit flag things that actively break behavior, or just unnecessary dead weight?

A more practical observation comes from Roei Shoustak: He tested and found that even before he typed a single word, Claude Code had already sent 61,652 tokens. Of that, `CLAUDE.md` and other instruction files took up 6,231 tokens, around 10%; while Claude Code's own built-in tools accounted for 23,818 tokens. That means the configuration files everyone loves to trim actually only make up a small fraction of the total.

But that doesn't mean you don't need to trim them. The most typical outdated habit is assuming the model needs you to teach it to "think step by step". Opus 5.5 doesn't need that. Rules like "only output code, no explanations" written for small context windows actually restrict the model's performance now.

### How to use it

Just enter this directly in a Claude Code session:

```

/claude-api prompt-audit

```

It will list which lines are redundant and which are suppressing the new model's capabilities, then you delete them manually. Don't expect a single command to fix everything for you — it only points out the issues.

Some developers have already run it and said they "have a lot to fix". Others have suggested adding a `/claude-api migrate` command next that updates both prompts and API calls in one go.

A user who wrote their `CLAUDE.md` in Vietnamese asked: Can it read rules written in other languages? They don't know much code, and their file is just full of plain sentences documenting what broke in the past.

The answer to that is still unknown. But one thing is certain: all the rules you added one by one in the past now need to be taken apart one by one.

If you haven't upgraded to Opus 5.5 yet, this command may not work for you. But if you have upgraded already, run it once, and delete those old "afraid of losing" habits.