A CLAUDE.md silent failure audit, because every other config yells when it breaks.

1. The thing that makes CLAUDE.md weird

Take a swing at this. Of every config file in your repo, which one fails the loudest?

tsconfig.json fails at compile time. Red squiggles in the editor, exit code 1 in CI, a refusal to ship a build. .eslintrc fails at save time. Husky blocks the commit. jest.config fails at PR time. A red check on GitHub, a Slack ping, a blocked merge. .env fails at boot. The app crashes, the deploy log spits a stack.

CLAUDE.md fails like none of those. It ships every byte to Claude on every turn, then says nothing. There is no log entry for "Claude saw line 47 and decided to skip it." There is no telemetry for "the prohibition fired but was overridden by an internal weight." There is no exit code. There is no red squiggle. The rule lives in the file, you assume it works, and the only evidence that it does not is the user-visible regression weeks later, plus the slow dawning thought of "wait, didn't I tell it not to do that?"

That is what makes a silent failure audit a different animal from every other lint pass in your toolchain. The whole point of the audit is to produce a signal where the runtime produces none.

2. CLAUDE.md vs every other config that yells

Sketch out the failure surface of one against the other:

The right column has six places the failure shows up. The left column has zero. The audit is the missing column.

3. The 7 silent-failure shapes the analyzer catches

The detector is one TypeScript file at src/lib/analyzer.ts, 322 lines, no network calls. It walks the file you paste and looks for these seven shapes. Each one is a discriminated case in the Finding union at line 5:

One sentence on each, with the file line where the detector lives so you can read the regex yourself:

4. Two detectors worth reading by hand

Most of the seven are obvious from the kind name. Two are not. The vague detector is just a banned-words list, but the contents of the list matter more than the regex. These 14 terms are the ones a senior engineer drops into a CLAUDE.md without noticing they have written nothing testable:

The cache_bust detector is the high-severity one, and it is also the one nobody else catches. A single line like "Today is 2026-05-20" in the first 20 lines invalidates the prompt cache for the entire file on every session:

The cost math: a 6,000-token CLAUDE.md with a cache hit costs about a tenth of a fresh read. One stale date in the first 20 lines, and every turn pays full input cost. Across a 30-turn session at Opus 4.7 input rates, that is the difference between roughly $0.27 and $2.70 spent on a file that has not changed.

5. A real audit on a 24-line CLAUDE.md

A representative file from a Next.js payments repo. 218 tokens, 24 lines, no obvious problems on a first read:

The audit output, top to bottom:

Nine findings, 74 tokens of recoverable waste per turn, and the file scores 5 of 12 on the Karpathy rubric. Every one of those nine lines would have shipped to Claude on every turn forever, producing no error and no warning, until somebody finally got curious. That is the gap the audit closes.

6. Five steps to run the audit yourself

7. What the audit does not catch (yet)

The 7 detectors are per-line, on the bytes of one file. That misses three other shapes of silent failure that live across files:

• CLAUDE.md vs settings.json contradictions. A "never run destructive shell commands" rule in your file, plus a permissive Bash allowlist in .claude/settings.json, leaves the rule visually present and operationally dead. The current analyzer does not cross-reference allowlists.
• Subagent inheritance. Subagents launched from your session do not inherit the parent CLAUDE.md. Any rule that mattered to the parent silently does not apply in the subagent context. See the sibling guide at /t/subagent-claude-md-inheritance for the per-subagent token math.
• Plugin overwrites. Some installable plugins write their own block into the bottom of CLAUDE.md and quietly clobber the previous content. The file you wrote yesterday is not the file that ships today. See /t/claude-md-plugin-overwrite-risk for the failure mode.

These three are on the analyzer roadmap. The current free tier handles the seven in-file shapes, which are the silent failures you can detect from the bytes alone, and which add up to most of the actual waste in a typical config.