Writing
Where AI autonomy breaks2026·07·11I checked 57 repos whose CLAUDE.md bans the any type. One in eight turns the rule off.Of 57 public repos whose CLAUDE.md or AGENTS.md tells the coding agent never to use the any type, only 46% back the ban with a hard lint error; one in eight ships a config that sets the rule off. A hand-verified census, and a deterministic experiment showing why your CLAUDE.md is load-bearing precisely where it's redundant.2026·07·07A Claude Code Skill with 86k Stars Says It Cuts Tokens 65%. It Cut My Bill 1%.A wildly popular skill promises to cut your AI coding tokens by 65% by making the model talk like a caveman. It works — the model really does write less. But those words are about 1% of a real session, so your bill — or your subscription quota — barely moves. Why token counts aren't costs, and where the savings actually hide.2026·07·06Verification Is All You NeedAn AI agent runs unattended exactly as far as you can verify its output cheaply, deterministically, and many times a night. That line doesn't split backend from frontend — it cuts diagonally: CRUD is cheap to check, the concurrent backend and the rendered frontend are not. Design your autonomy around the gradient.2026·06·05Passed every test. Still took down prod.It passed every test and still took down prod — the bug wasn't a bad line, it was a state nobody enumerated. An LLM writes half your code now and nobody reads every line, so the state space is the only contract you can still enforce. A discriminated union deletes the bad states outright; a database constraint holds where the type can't reach. The two cheapest cuts — and the ones to make first.2026·06·05The invoice was dead. Then it took your money.A voided invoice that still captured a payment — green tests, red ledger, and a bug that lived in the ordering, not in any single line. When you can't delete a bad state, you verify it: state machines, covering arrays, model checking, and a spec that's now cheap to write — the verification layer, cheapest first.