Routing a 194,492-character CLAUDE.md by failure cost left 47 rules resident

The saving came from the load trigger. CLAUDE.md is resident, so every character in it loads on every message ^[2]. A skill loads when the model judges its description relevant, and the author is explicit that this is a judgment ^[8]. A rule file under `.claude/rules/` with `paths:` frontmatter loads by glob match on the files you touch, with no model judgment involved ^[7].

`@import` is not worth doing for the same reason. Imports expand inline: you get a tidier file and identical context cost ^[6]. Deleting was ruled out too, because you cannot know which lines are stale without re-reading the code each one describes ^[28].

Importance could not sort the file, since all of it was important ^[24]. The author's test is a failure question instead: "If this rule failed to load exactly when it mattered, how bad is that?" ^[10]. A styling trap costs time; a security rule ships the bug it exists to prevent ^[11]. So all 26 security rules stayed resident, along with 21 architectural invariants and the commands ^[12], while fees, booking, notifications, the mobile apps, the rendered pages, test practices and the known-gaps list moved out ^[13]. The two resident groups add to 47, the number of rules the write-up says the roughly 200-line file states ^[25].

Compressing 47 unverified rules into one-liners is the move the piece argues hardest against, because the output would be 47 confident-sounding claims nobody checked, looking freshly reviewed ^[15]. The author quotes his own repo's guidance, written after getting burned: "A stale claim that gets edited without being checked is laundered, not corrected. Re-read the code the line describes, or leave the line alone." ^[16]. The workaround is mechanical. Lift the rule's existing opening sentence verbatim and push the reasoning one hop away ^[17]. The Stripe entry keeps one sentence in CLAUDE.md, saying the card never touches this server, that a SetupIntent hands the browser a client secret and Stripe's own SDK collects the number, and that this is why no endpoint here takes one ^[19], and parks the other 400 words in `hard-rules/references/security.md` ^[18].

For the 91% figure to transfer, your rules have to already be written with the rule as the first sentence; the author says his were ^[17]. The disclaimer says the scripts were generated with AI assistance and the workflow was tested end to end on one repository, with all quoted numbers from that run ^[22]. On size, the old file was about 4.9 times the point at which Claude Code warns about a memory file ^[3], which puts that threshold near 39,700 characters and leaves the 17,283-character replacement at roughly 44% of it ^[26].

Verification is metered. The harness launches real headless Claude Code sessions, one per prompt, each consuming API quota and taking several minutes, and the author suggests starting with two or three prompts before running twenty ^[20]. It also rewrites CLAUDE.md in place and creates files under `.claude/`, so commit first ^[21]. The four words lost in the cut were all from headings deleted on purpose ^[4]. That count is a measure of text, not of behaviour. The published text explains how to run the harness but does not include its results ^[29], though the author calls the measurement "the most useful part and the part nobody writes about" ^[23].