Claude Code adds prompt-audit to catch stale prompts

Claude Code 2.1.221 adds a prompt-audit command that scans system prompts and MCP tool descriptions for patterns written for older model generations, per the Claude Code changelog. That's an invisible tax on every request: prompts accumulate few-shot blocks and defensive instructions that made sense two generations back. Now they just burn cached tokens or steer the model in ways nobody intended.

The command is a mechanical first pass, not a replacement for hand-editing a long-lived prompt, according to the changelog. It flags verbose few-shot examples, defensive formatting rules, and reasoning nudges that used to earn their keep. Now they just sit there, unread by anyone but the model.

It doesn't rewrite anything. Someone still has to decide what survives the cut.

That tracks with how these prompts get written. Someone adds a hedge or an example for one failure mode, and nobody ever removes it once the model stops needing the crutch. A scanner that flags the stale parts before a human rewrites by hand is a smaller ask than it sounds, and a useful one.

What matters here is the admission buried in the feature: prompt-engineering advice has an expiration date. Every long-lived system prompt is carrying dead weight from a model generation or two back, and most teams have no idea how much.