Skills
Audit your prompts and tool descriptions for patterns written for older models: new `prompt-audit` subcommand in the claude-api skill
Claude Code 2.1.221 adds a `prompt-audit` subcommand to the `claude-api` skill that scans prompts and tool descriptions for patterns written for older model generations. This targets a real and largely invisible tax: system prompts and MCP tool descriptions accumulate scaffolding — verbose few-shot blocks, defensive formatting instructions, reasoning nudges — that earned its keep two model generations ago and now just burns cached input tokens or actively steers behavior. It pairs directly with the finding that misleading architectural hints cost more than irrelevant prose, and gives you a mechanical first pass before hand-editing a long-lived prompt.
↳ Follow the thread