Your CLAUDE.md gets more expensive every week. Nobody is doing it on purpose.

Every guide on this topic shows you what your CLAUDE.md costs in one moment. That number is a y-coordinate. The bill is a slope. Two files at 6,000 tokens today look identical to an audit; one was 5,800 last week and one was 2,100. The first is stable, the second is on track to be 12,000 by July. The audit cannot tell you which is which.

This is what we mean by drift. It is the only reason a senior engineer's file looks fine in week one and unaffordable in week eight. None of the commits in between were wrong.

1. The analyzer reports one instant. Drift hides between instants.

The ccmd analyzer is a pure function. You paste a CLAUDE.md, it returns one snapshot. The two lines that produce the cost number are this:

Three things to notice. First, estimatedTokensFireEveryTurn = totalTokens. The analyzer assumes the entire file fires on every turn, which is what Claude Code actually does with project-level CLAUDE.md. Second, token estimation is Math.ceil(text.length / 4), the same chars-divided-by-four heuristic Anthropic's own tokenizer-counting CLI examples use. Third, the cost is per long-running session of 30 turns at Opus 4.7 input rates.

That is all you need to measure drift. Run the same function on two CLAUDE.mds from two SHAs and the difference is your weekly delta. The analyzer never does that because it does not have your git history; you do.

2. Eight weeks, six commits, 3x cost

This is the actual shape we keep seeing in repos. Six commits across eight weeks that nobody would block in review. The number on the right is the per-turn token floor at that moment.

Run the audit in week 0 and the file is fine. Run it in week 8 and the file is bad. Run it in any week in between and the finding count climbs by one or two. None of the snapshots look alarming. The trajectory does.

3. The shape of drift, on your repo, in two commands

Save this to scripts/claudemd-drift.sh and run it from your repo root. It is the entire drift detector, written from the ccmd analyzer's own constants. No network calls, no signup.

Here is what it looks like on a real repo we audited last week:

312 inserted lines, 7 deleted. That 44-to-1 ratio is the fingerprint of drift. A file that grows the way code grows (delete-rewrite-refactor) does not look like that. A file that grows the way a log file grows does. CLAUDE.md is a log file unless somebody is paid to compact it.

4. Snapshot vs slope, side by side

Most existing playbooks on this topic stop at the snapshot. The snapshot finds the one bad line, sometimes finds the duplicate, and is silent about the trajectory. Here is what falls into each column:

Both columns are real and both have a place. The snapshot tells you what to delete today. The slope tells you whether the file will be back at this size in six weeks regardless of what you delete.

5. The one-way function: why almost nobody deletes a CLAUDE.md line

The structural reason drift is monotonic upward is human, not technical. When something breaks, the cost of adding a rule is felt (we got burned). The benefit is immediate and named (the rule will stop this from happening again). When a rule becomes obsolete, neither the cost nor the benefit of deletion is felt: nothing ever notifies you that a rule is stale, no test fails when a rule is duplicated, no PR is blocked when a section becomes irrelevant. So the add path is high-signal and the delete path is silent. Over weeks, that asymmetry is sufficient to explain the slope.

The fix is to make the delete path loud. The ccmd analyzer already does this on the snapshot side: the duplicate finding flags a rule that already exists earlier in the file, the vague finding flags adjectives like "appropriate" that have no testable meaning, the bloat finding flags any line over 28 words. Those are deletion prompts. The paid drift monitor sends them weekly so the prompt arrives on the day the rule landed, not eight weeks later.

6. What to actually do this afternoon

1. Paste your current CLAUDE.md into ccmd.dev. Record the totalTokens number. That is your y-coordinate today.
2. Run git log --since='8 weeks ago' --shortstat -- CLAUDE.md. Sum the insertions and deletions. If insertions are more than 10x deletions, drift is happening at structural rate.
3. Use the eight-week-old SHA from git rev-list to recover the previous CLAUDE.md, paste it into the analyzer, and record that totalTokens. Subtract. Divide by eight. That is your tokens-per-week growth rate.
4. Set a per-PR token budget for CLAUDE.md. We use 4,000 tokens on most projects. Any PR that pushes the file over needs to name the existing rule it replaces or move the new content into a scoped sub-directory CLAUDE.md (see the multi-file token budget guide) or a Claude Skill that loads on demand.
5. If you have a single line near the top that contains a date, the word "today", "this session", or "right now", delete it first. That is the cache_bust finding and it multiplies the cost of every drift point by roughly 10x via the prompt cache. See the related weekly quota burn walkthrough.