The per-line cost of CLAUDE.md, and why it is the only number worth editing on

Why per-line, not per-file

The whole-file per-turn cost is famously a flat number: your CLAUDE.md fires in full on every API call, so per-turn dollars are token_count x rate / 1M. That is a number to be aware of. It is not a number you act on, because there is no button labeled "delete some bytes" in the middle of a file.

What you act on is one line at a time. So the useful unit is what that one line costs while it sits in the file. Total cost is the sum of per-line costs; if you cut a line, its tokens leave the sum on every turn, every session, from now until the line goes back in. The per-line dollar is the lever; the per-file dollar is the dial it moves.

Concretely: same chars/4 estimator, same input rate, scoped to one line:

The per-line ladder, on a real 35-line CLAUDE.md

Below is the actual sample file the analyzer ships with (lines 288-322 of src/lib/analyzer.ts), one row per non-blank line, ranked by cost. Numbers use chars/4, Opus 4.7 input rate $5/M, 30-turn session, and a team of 5 engineers running 30 sessions per engineer per week (so the per-team-year column is per-turn x 234,000).

Total per-team-year for the visible 20 rows: about $322. That is the cost of this sample CLAUDE.md continuing to exist, unedited, across one year of one 5-engineer team. The whole-file per-turn cost says "your CLAUDE.md costs X per session." The per-line ladder says where X comes from.

The 425x cost spread, in one file

The cheapest line in the sample (## Never, line 300) costs $0.00001 per turn. The most expensive (the run-on on line 308) costs $0.000425. Same file, same firing model, same rate, 42.5x cost spread on per-turn basis. Scale to per-team-year and the spread is the difference between $2.34 and $99.45 per year.

The lesson is not that headers are free. Headers are almost free, which is fine. The lesson is that the long lines pay for everything else. Three or four sentences like the stack run-on at line 308 will outweigh the entire rest of the file. When you audit a CLAUDE.md for cost, you are looking for those three lines.

What our analyzer does is rank exactly this: each line goes through the chars/4 estimator, gets a per-turn token count, and gets cross-checked against seven deadweight categories (bloat, vague, aspirational, missing_why, duplicate, conflict, cache_bust) so the ranked list also tells you which of the expensive lines are doing real work and which aren't.

The cut-order rule

With per-line numbers in hand, the cut order is mechanical. Read each row top-to-bottom in the ladder and apply this filter:

1. If the line is flagged duplicate, cut it. The rule still exists on its first appearance; you cannot regress behavior by removing a second copy. In the sample above, line 316 is the cheapest cut in the entire file.
2. If the line is flagged cache_bust and is in the first 20 lines, fix it. Either move the volatile string below idx 20 or delete it. This is the one finding that costs you ~10x the listed per-line dollar across the whole file, not just the line itself.
3. If the line is over 28 words (bloat), split it. Same word count, half the lines, no token saving. But the model treats each line as one signal, so splitting raises the chance the back half actually fires.
4. If the line is vague or aspirational with no "unless" and no "why", rewrite or cut. A line that the agent silently ignores still pays per turn. The team-year column tells you which of these are worth the rewrite effort and which to just delete.
5. Move the expensive lines that survive (1)-(4) into a skill.A rule that fires only when its description matches stops counting against per-turn cost entirely. That is how the per-line ladder shrinks even when nothing got "cut."

Each step reduces the sum the per-turn cost is computed from. The order matters because step 1 is risk-free, step 2 is the largest per-edit return, and step 5 keeps content alive while taking it off the every-turn surface.

The sample file, so you can verify any row

Every row in the ladder above comes from the file below, which lives at src/lib/analyzer.ts lines 288-322. To verify any row: count the line's characters, divide by 4, multiply by 5, divide by 1,000,000 for $/turn. The token column rounds up by Math.ceil, which is also what analyzer.ts:38 does.

What the per-line number does not include

Three honest qualifiers, because the ladder above is the floor, not the ceiling:

• Opus 4.7's new tokenizer is heavier than chars/4 by up to 35% on code, punctuation, and non-English text. So the per-line tokens are a low estimate. A line the ladder calls 15 tokens may bill closer to 20 on Opus 4.7. The relative ranking is correct; the absolute dollars are a floor.
• Output tokens are billed separately and are not part of the per-line cost. The per-line ladder is input-only, which is what CLAUDE.md is. Output cost depends on what the agent ends up doing in that turn, not on the configuration.
• The per-line ladder ignores layered files. If your project loads both ~/.claude/CLAUDE.md (global) and ./CLAUDE.md (project), the per-turn cost is the sum of both files' per-line costs.

None of these change the cut-order rule. The most expensive lines are still the most expensive lines, and cutting them still saves the same proportion of the bill.