A CLAUDE.md context-burn audit, the seven things to actually grade.

Most posts that show up under "audit your CLAUDE.md" say the same thing. Keep the file under 500 words. Remove instructions you added three months ago. Every line gets sent on every turn. That is all true and none of it tells you which lines.

An actual audit fires concrete checks against the file you already wrote and points at the lines. This page walks through the seven checks ccmd runs, the exact threshold or regex each one fires on, and the order to fix them in. The whole thing takes 220 ms in your browser.

1. The input

A representative CLAUDE.md from a real-feeling Next.js payments repo. 213 tokens, 22 lines. Looks fine on first read. Scores 5 of 12 on the rubric and ships 8 findings.

2. What the audit says

Same file, run through analyzeConfig() in src/lib/analyzer.ts. The UI on ccmd.dev renders the same data as a richer panel; the shape is below.

Eight findings on 22 lines. Read top to bottom: line 3 is the single highest-severity finding on the page and the cheapest fix. Lines 7, 8, and 9 are absolute-word and vague-term cleanups. Lines 12 and 13 are prohibitions with no Why. Line 17 is a stack paragraph that wants to be three short rules. Line 22 is a copy-paste of line 7. The seven check categories on the right column come next.

3. The seven checks the audit runs

Each one is a pure function in the same file. Thresholds and regex are below; copy them, adapt them, port them. The whole module is under 300 lines.

4. The math under the audit

The audit assumes the file fires every turn (it does), 30 turns per long-running session (a heuristic from the docs), and the published Opus 4.7 input rate. These three constants give you a defensible dollar number on the file you wrote.

For the 213-token sample above, the dollar math reads: 213 tokens × 30 turns × $5 per million = $0.032. Trivial in isolation. Now scale the file to a realistic 6,000-token CLAUDE.md and the same math comes out to $0.90 per session, or about $27 per month at one long session a day. With a cache hit that drops to about $2.70 per month. Without cache (because one ISO date at line 3 invalidated it), back to $27. The cache_bust finding moves the cost by 10x.

5. What the other "audit your CLAUDE.md" pages do

We read the top results that currently rank for this topic. They agree on the diagnosis (sloppy CLAUDE.md taxes every turn). They stop short of the audit.

6. The order to fix in

Findings have different blast radius. Cheapest-impact-first beats cheapest-effort-first. For the sample above the order is:

1. cache_bust on L3. Strip the date line or move it to the bottom. One delete, recovers the full prompt-cache discount for the rest of the session.
2. duplicate on L22. Delete the second copy. One line, zero risk.
3. missing_why on L12 and L13. Append a Why: with the past incident. One line each. "Never use any. Why: silent runtime error in checkout, 2026-03, 4 hours to trace."
4. bloat on L17. Split the 31-word stack paragraph into three short rules: framework, payments boundary, migration rule. The second half of the original was being ignored anyway.
5. vague + aspirational on L7-9. Replace "always handle edge cases appropriately" with "cover the four cases in cases.test.ts". Replace "always think carefully" with a tests-must-pass line. The absolute-word findings often go away when the line gets concrete.

After all five passes the file scores 9 of 12 on the rubric, drops to roughly 145 tokens, and the prompt cache stays warm. The token saving is real but secondary. The larger win is that the agent stops ignoring half of what is in the file.