After reading the full history of thinking machines: Vote below, then tell us in the comments: which part of the timeline surprised you the most?

Today I was working between Opus and Fable when I noticed our CLAUDE.md had crept to 75 lines, over the recommended 50 outlined in Lesson 3.3 (Mistake 1: making the CLAUDE.md too long). So I asked Claude to review it. The useful discovery wasn't the trim. It was that line count is only a smoke alarm and the real test is whether each line actually changes Claude's behavior. We got from 75 to 55 with zero instruction loss, and this is what Claude shared with me afterwards: Trim done — 75 → 55 lines, zero instruction loss. Backstory is now a pointer to CONTEXT.md, Status delegates to HANDOFF (which I verified holds everything the deleted lines said — due date, schema, locked status set), and the archive note is compressed to its one live fact. Now your real question — what to look for going forward so you're helping, not just counting. Your line-count instinct was a good smoke alarm: it got us to look. Here's the sharper toolkit, in order of usefulness: 1. Ask the function question, not the length question. Scan any always-loaded file and ask of each line: "does this change what Claude does?" Instructions and routing earn their place. History, narrative, and state don't — they belong in CONTEXT, HANDOFF, or git. That one question did all the work in today's trim. 2. Hunt duplication — one home per fact. The Status section was the genuine hazard today, not the length. Anytime you see the same fact in two files, ask "which is canonical?" and make the other a pointer. Drift between copies is the failure that actually burns us. 3. Watch for time-bombs in always-loaded files. Dates, "due", "currently", "Status:" — anything time-bound in CLAUDE.md will go stale, because nothing forces an update there. State belongs in HANDOFF, which gets overwritten by design. 4. The best prompt you can give me, occasionally: "Which CLAUDE.md lines did you actually use this session?" That's ground truth no heuristic matches — lines I never act on across several sessions are candidates to demote. 5. Trust the tripwires we've now written (folder >10 docs → room map; state never in CLAUDE.md). You don't need to re-derive these — they're in the file, so I'm bound by them even when neither of us is thinking about it.