Your CLAUDE.md token budget per turn is one ceiling, not an allowance.

The phrase "per turn" misleads you

A budget, in normal usage, is something you spend down and get back. A monthly budget refills monthly. A per-request quota refills per request. So "token budget per turn" reads like a discretionary pool: some tokens you choose to spend on CLAUDE.md this turn, some you save. That mental model is wrong, and it is the reason people keep tuning the wrong thing.

CLAUDE.md does not get a per-turn allowance. It is a standing reservation. The file's entire token count sits in the system prompt on every single turn, unchanged, whether the turn is a one-line question or a 40-tool refactor. You do not decide per turn how much of it to load. You decided that once, when you saved the file at whatever size it is.

The proof: the analyzer never models a per-turn fraction

Our analyzer is a pure function with no network calls. When it computes the per-turn token figure, there is no fraction anywhere in the code. It estimates total tokens as characters over 4, then sets the per-turn number equal to that total. That is the whole calculation.

This matters because it tells you exactly which number to budget. There is no "portion that fires" to tune down. There is no partial load. The per-turn token cost of CLAUDE.md is a property of the file, full stop, so the budget is a property of the file too: one ceiling on totalTokens. Anthropic's own cost guidance says the same thing in plain English: your CLAUDE.md is "loaded into context at session start" and its tokens "are present even when you're doing unrelated work."

Setting the number

A budget you never picked is not a budget. So pick one. The cleanest anchor is Anthropic's own line in the Claude Code cost docs: "Aim to keep CLAUDE.md under 200 lines by including only essentials." That is a line ceiling. To turn it into a token budget, count characters and divide by 4, the same estimate the analyzer uses.

A 200-line file averaging 45 to 55 characters per line is about 9,000 to 11,000 characters, which is roughly 2,250 to 2,750 tokens. Sparse files with short directives land lower; dense prose lands higher. Here is the conversion as three working tiers. Pick the row that matches your file's job, then treat its token number as your ceiling.

Line numbers anchor to Anthropic's 200-line guidance; token numbers are the chars/4 conversion at typical CLAUDE.md line density. The monorepo row assumes the root file plus what nested files add; everything in the firing prefix counts against one combined ceiling.

A budget is necessary, not sufficient

Here is the part most cost guides skip. Hitting the budget does not make the file good. We ran our analyzer on the SAMPLE_CLAUDE_MD constant that ships in src/lib/analyzer.ts, a deliberately bad example file. It is 280 tokens across 35 lines: comfortably inside even the tightest budget row above. And it still fails on every quality axis.

The 18 findings break down as 10 aspirational absolutes with no escape clause, 5 vague terms with no testable meaning, and 3 prohibitions with no stated reason. None of them carry a tokenSavings number, which is why recoverable tokens is 0. The fix is not deletion, it is rewriting. A budget caps what the file costs you per turn. It does nothing about whether the file works. You need both: a size ceiling and a quality pass.

Enforcing the ceiling

Once you have a number, enforcement is a comparison: file token count versus ceiling. Paste your file into the analyzer on ccmd.dev and it reports totalTokens and estimatedTokensFireEveryTurn (always equal). It runs in your browser, no upload, no signup. Here is what a check looks like on a file that is over a 2,500-token budget:

When you are over, the per-turn cost is not the first thing to cut. Two cheaper moves come before deleting content. First, move conditional rules off the every-turn surface: a rule that only matters during database migrations or PR reviews does not belong in a file that fires every turn. Anthropic's docs say it plainly, "move instructions from CLAUDE.md to skills", because skills load on-demand and stop counting toward the per-turn budget entirely. Second, kill duplicates and split bloated lines, the bytes inside your budget producing nothing. Trimming actual content is the third lever, not the first.

A budget set once does not stay met. Every commit that touches CLAUDE.md is a chance to drift back over the ceiling, and the file tends to grow because adding a rule is easy and removing one feels risky. The free analyzer is a one-shot check. The paid tier ($9 to $19 per month solo, $49 for a team) watches the file over time, sends a weekly drift email, and posts a diff comment on any pull request that pushes the file back over your budget.