Ulrick

THIS IS ACTUAL F*CKING GOLD Anthropic just shared one of the best tricks for Claude Code. The idea: stop asking for an answer. Give it a stop condition and let one agent loop until it hits it. Same codebase, two ways: one-shot got 61% and guessed the rest. The loop ran 43 times and came back with 388 of 388 rules verified. You just launch Claude Code like this: claude --model opus --loop "map legacy/ until 100% verified" here's the split: → you write one prompt and then do nothing, zero edits for the entire run → REA loops read → extract → verify → fix on its own → every rule it extracts gets replayed against real prod logs before it counts → the loop only stops when all 388 rules pass, not when the model feels done the part everyone gets wrong: they accept the first pass. a single pass over 16 services gives you a confident document where most of it was never checked against anything. here rules.md grows every loop. 388 rules across auth, billing, payments, checkout, admin and the rest, each one marked verified only after it survives a replay. three rules came back wrong. the agent re-read the source, fixed them, and looped again instead of shipping them. one prompt from you, 43 loops from it, 0 edits by you. Opus drives. REA reads and verifies. The loop decides when it is done. JEV engineering is the same move one layer down: the forks that need no thinker (which file next, re-loop or stop, retry or fail) go to JEV in milliseconds, and the model only sees the ones that actually split the full tree > Opus 5.5 runs the session with a stop condition, not a task > REA loops read, extract, verify and fix on every pass > every extracted rule is replayed against prod logs before it counts as verified > any mismatch sends that module back for a re-read instead of shipping > rules.md grows each loop until the stop condition is met > JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the main session on Opus 5.5 and give every long task a stop condition, never a single pass. 2. Loop read, extract, verify and fix until the condition is met. 3. Every extracted rule must be replayed against real logs or tests before it counts as verified. 4. Any mismatch triggers a re-read of that module, never a patched guess. 5. Add one rule to CLAUDE.md: unverified findings are never written into the final document. 6. Route deterministic forks (next file, re-loop or stop, retries) through JEV. Show me every change as a diff first. No edits until I say go." ↳ code.claude.com/docs/en/mode… save it, you'll use this on every legacy repo from now on.