← ClaudeAtlas

auditlisted

Show what an agent loads on every single request and what is wrong with it — resolve CLAUDE.md, AGENTS.md and their imports into one hot-context inventory, price it in lines and estimated tokens, and flag oversized files, broken imports, references to files that no longer exist and instructions duplicated across files. Use when the user asks what is in their context, why their agent seems to ignore instructions, whether CLAUDE.md has grown too big, wants to review or clean up their instruction files, says "CLAUDE.md'ye bak", "bağlamımda ne var", "audit my instructions", "is my CLAUDE.md too long", or runs /gardener:audit. Not for deciding what to remove — that is gardener:prune. Not for checking whether rules are obeyed — that is gardener:compliance.
hailneed/gardener · ★ 0 · Code & Development · score 72
Install: claude install-skill hailneed/gardener
# Hot context audit (audit) Instruction files are the most expensive text in a repository: every line is paid for on every request, forever. This shows what is actually loaded, what it costs and what is broken. Local only, no quota. **Language rule: write every user-facing output in the language the user is speaking with you.** **Honesty rule: token counts are a bytes-over-four estimate, not a real tokenizer. Say "estimated" every time you quote one; never present it as an exact figure.** ## Step 1 — Inventory ``` node "${CLAUDE_PLUGIN_ROOT}/scripts/gardener.mjs" --audit --repo "<repo>" --md ``` Default `--repo` is the current directory. Read `${CLAUDE_PLUGIN_ROOT}/skills/audit/references/checks.md` before triaging — it explains what each check means and when it is safe to ignore. The report resolves `@path` imports transitively, so the file list is what actually loads — not what the top-level file appears to say. A file marked ⚠ failed to resolve. ## Step 2 — Read the cost table first The findings matter less than the first table. Three questions, in order: 1. **What is the total?** Under ~4k estimated tokens is comfortable. Past ~8k the agent starts every request with a large fraction of its attention already spent. 2. **Which file dominates?** Usually one file is 70–90% of the total. That is where the work is; everything else is rounding. 3. **Is anything loading that you forgot about?** Import chains are the usual surprise — a global file pulling in thre