CLAUDE.md check: what it costs per session and what does not belong in it
CLAUDE.md is read on every session and re-read on every message. This page tells you what that costs and which lines are the usual culprits: statuses written as if they were rules, and step-by-step procedures that should live in a skill and load only when needed. Paste the file; nothing is uploaded.
What the check does and does not do
It counts tokens by a rough ratio (about 3.6 characters per token for mixed English and code; Cyrillic runs closer to 2.5, so Russian files are underestimated). It flags lines by pattern: dates, "currently", version numbers, work-in-progress markers, and step-like wording. Patterns miss meaning in both directions, so treat the lists as places to look, not as a verdict.
The rule behind it, from running an agent whose CLAUDE.md is read four times a day: rules in the file that always loads, and keep it short; procedures in skills that load on demand; state in files that scripts write. The reasoning and the test for each file is in what Claude Code reads on every session.
Privacy
Parsed by the JavaScript on this page. Nothing is sent anywhere; the page has no network calls.