Blog
Numbers we counted on real machines, and the things that broke while we were building. Measurements first. Opinions after, if at all.
September 23, 2026
I grew my agent instruction file from 90KB to 147KB. Violations did not drop.48 days on one machine. 94.4% of tokens were re-reads of the same instructions. Across 24 tracked rules, the 8 moved into code gates recurred 0%; the 16 left as document clauses recurred 100%.
September 15, 2026
Shopify's "split brain" is not a filename dispute. It is drift, and one standard file will not end it.Shopify's CEO threatened to ban Claude Code over AGENTS.md vs CLAUDE.md. The failure he described is instruction-file drift. A standard filename removes one multiplier, not the mechanism. Two practices that do.
August 19, 2026
The ledger said 763 assets. The platforms said about 400.An asset database counted 763 items collected from 21 AI platforms. Counting by hand found roughly 400 real ones. The three ways a sync ledger lies: ghosts, misses, and false positives.
August 19, 2026
24 of the 30 duplicate groups were not duplicatesGrouping by title found 30 duplicate groups. Content hashing dissolved 24: same-title different assets, test probes, and empty rows from one bad batch. What actually identifies a duplicate.
August 17, 2026
AGENTS.md was supposed to end the sprawl. I counted the files on one laptop.One laptop: 111 AGENTS.md files with 19 distinct contents, 38 CLAUDE.md with 19, 3,073 SKILL.md with 887. Where each agent reads its instructions, and why copies diverge.
August 13, 2026
Why your AI agent instruction files drift across machinesYou write CLAUDE.md once. Six months later there are eighteen of them on your laptop and you cannot say which one is right. Why it happens and why the obvious fix does not fix it.
August 11, 2026
Generating your agent rules from one file does not stop them driftingWrite AGENTS.md, generate CLAUDE.md and .cursor/rules from it. It works on the day you run it. Here is the failure mode that shows up afterwards, measured on a real machine.
August 11, 2026
I found 18 versions of the same CLAUDE.md on one laptopClaude Code, Cursor and Copilot each read a file before they act. Those files get copied, then edited. A scan of one working machine, and what the spread actually looked like.
Connect one workspace and see every skill, rule, and memory your team has in play — in about ten minutes.
No credit card. Works with what you already run.