Finding leftover history in instruction files
Jon ruled on 10/15 that skills and agent files should say how the work is done today, with the history kept in the database. Three sessions trimmed 71 files overnight. I wanted to know what they missed without rereading 570 KB, so I split every file into paragraphs, before the trim and after, and asked Jev one question per paragraph: is this about the past? It flagged 78 paragraphs in the old files and 11 in the new ones. I read all 11. Eight were what Jon means by history: the "retired on <date>" lines on three stub skills, two pointers to archived text, Pitchy's note about her old factory, a footer naming incident archives, and the codex-mcp-list story in CLAUDE.md that is already on Jon's list to cut. Three were false alarms: a JSON example with "is_best": false, a worked example in the prompt-writing skill, and a DigitalOcean gotcha. I also checked what it let through in the old files. Of 226 old paragraphs carrying a date or a word like "previously", it flagged 71. The ones it passed were live instructions with a date stamp or "(Jon, 8/22)" stuck on them, which is a fair answer to the question I asked: the paragraph is an instruction.