Boban Alempijevic

THIS IS A PURE F*CKING GOLDMINE Anthropic just shared one of the best tricks for Claude Code The idea: run /goal in auto mode, let Opus 5.5 frame the bounds, stream Haiku 5.5 evaluations at 340 tok/s, let Sonnet 5.5 build, and keep Fable 5.1 on call as the dedicated adversary: Opus-level judgment on every big milestone, without paying Opus rates for every routine turn: Claude Code runs the loop in auto mode and Opus only signs off on the verified PR You just launch Claude Code like this: claude --model opus here's how it splits: → Opus 5.5 runs the main session, defines the verifiable finish line, and locks bounds → Sonnet 5.5 acts as the builder, applying code edits and running pytest suites in isolated turns → Haiku 5.5 streams transcript evaluations at 340 tok/s after every single turn → Fable 5.1 stays on call as dedicated adversary to audit goals, break loops, and review diffs → you stop burning Opus context on dozens of intermediate autonomous iterations Fable 5.1 only steps in at three moments: → before the loop starts: audits the /goal condition to ensure criteria are objectively verifiable → when a test fails twice: stops symptom patching and redirects the implementation approach → before calling it done: cold adversarial review of diffs and test suites before Opus sign-off Opus bounds · Sonnet builds · Haiku evaluates · Fable audits JEV engineering is the same move one layer down: the forks that need no thinker (which tool, path validation, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split the full tree > Opus 5.5 runs the main session and locks the verifiable goal ├─ builder: Sonnet 5.5 refactors code and runs pytest in isolated turns ├─ evaluator: Haiku 5.5 evaluates transcript at 340 tok/s after every turn ├─ governor: Haiku 5.5 tracks turn budget and halts the loop on Met verdict ├─ Fable 5.1: adversary on call, audits goals, loops, and final PR diffs └─ JEV: routes micro-forks and deterministic checks 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 with full reasoning depth 2. When running /goal, let Opus frame objectively verifiable completion criteria before starting the loop 3. Stream stop hook transcript evaluations on Haiku 5.5 at 340 tok/s and let Sonnet 5.5 build and test 4. Keep Fable 5.1 on call as dedicated adversary to audit goal conditions, double-failure loops, and PR diffs 5. Route fast deterministic forks through JEV, show all changes as diffs first. No edits until I say go" ↳ code.claude.com/docs/en/goal save it, you'll use this every single day