The Lean CLAUDE.md Sheet — What to Delete From Your CLAUDE.md After Anthropic Cut 80% of Their Own System Prompt
The companion to the full video. On July 24, 2026 Anthropic published 'The new rules of context engineering for Claude 5 generation models' and said they had removed over 80% of Claude Code's system prompt for Opus 5 and Fable 5 with no measurable loss on their coding evaluations — then shipped the rules they used inside the /doctor command. This sheet is those rules on one page. INSIDE: (1) THE ONE TEST that decides every line, quoted verbatim from the shipped /doctor instructions — 'could a session working in this repo reconstruct this by reading the code? If yes, cut it. If no, keep it' — plus the second rule that matters just as much: when unsure, keep it. (2) ANTHROPIC'S OWN CUT LIST — the seven categories a fresh session rebuilds for itself: directory trees, tech-stack lists, standard build and test commands, schemas copied out of your own source, README-style architecture tours, generic best practice, and anything a lint config or CI already enforces mechanically. (3) THE KEEP LIST — gotchas and failure contracts, design rationale, conventions that DIFFER from the framework default, safety prohibitions, repo etiquette, domain glossaries, non-guessable commands, and pointers to context elsewhere. (4) THE FOUR DESTINATIONS and exactly when each one loads: the root CLAUDE.md (always), a subdirectory CLAUDE.md (only in that folder), a skill at .claude/skills/<name>/SKILL.md (only when invoked, with just the one-line description resident), and .claude/rules/*.md scoped by a `paths` value in the frontmatter — plus the one hard exception, that a 'never do X' rule must never be moved into a lazily loaded file. (5) A REAL BEFORE AND AFTER — a demo repo was built for the video with a CLAUDE.md written in standard 2025 style, 174 lines, and the real /doctor was run against it: about 1,585 estimated tokens loaded every session, down to about 558 after the trim and one migration. Including the worked example that teaches the whole boundary — the style bullet /doctor refused to delete, because it deliberately contradicts the framework default. HONESTY: all token figures are estimates (characters ÷ 4) — run /context for the exact live measurement. The 65% reduction is that demo repo's number, on a file written deliberately bloated; yours will differ. What was measured is token cost, not output quality. Independent field sheet from Hyperautomation Labs — not affiliated with Anthropic.
Free. No spam. Unsubscribe anytime.