auditlisted

Audit a project's Claude Code instruction surface (CLAUDE.md, .claude/rules, agents, skills, settings) for instructions written for older Claude models that current models no longer need or that now backfire, and propose a diff. Run it after a model upgrade, or when turns feel slow, verbose or over-delegated.
inbrace-tech/claude-skills · ★ 1 · AI & Automation · score 77

Install: claude install-skill inbrace-tech/claude-skills

# Audit Claude Code configuration for outdated instructions Prompts and rules written for an older model keep running after every upgrade. Some become dead weight, and some now work against the model: an instruction that fixed a weakness of Claude Opus 4.8 can make Claude Opus 5 or 5.5 slower, more expensive or less accurate. This skill finds those instructions and proposes what to change. ## Scope and defaults - **Path:** the argument, or the current project root when none is given. - **Target model:** `--target <model>`, or else the model this session runs on. State the target in the report. - **Apply:** propose only. Edit files only when the invocation carries `--apply` or the user asks for it after reading the report. Do not stop to ask about scope. State your assumptions at the top of the report instead. ## Step 1: Inventory List every file that shapes model behavior, and read each one whole: - `CLAUDE.md`, `CLAUDE.local.md`, `AGENTS.md`, and every file they import with `@path`. - `.claude/rules/**/*.md`, `.claude/agents/*.md`, `.claude/skills/*/SKILL.md`, `.claude/commands/*.md`. - `.claude/settings.json` and `.claude/settings.local.json`: look at `env`, `model`, and hook commands that inject text. - Any application code in the repository that calls the Claude API: search for `thinking`, `budget_tokens`, `effort`, `tool_choice`. `rg` skips dot-directories when it walks from `.`, so name `.claude` explicitly or pass `--hidden`. ## Step 2: Scan for these patterns