Opus 5.5, Haiku 5.5 and Jev three-agent chain builds a project valued at $43,000
Opus splits tasks, Haiku executes them one by one, and Jev only checks each against its "done when" line and returns pass, fix or reject; nothing merges without Jev's pass, with a Claude Code config included.
OPUS 5.5 + HAIKU 5.5 + JEV BUILT A PROJECT FOR ME THAT JUST GOT VALUED AT $43,000. NONE OF THEM COULD'VE DONE IT ALONE i read the guide and it clicked. this thing is actually powerful → Opus 5.5 = the architect. reads the brief once, splits the project into small tasks. each task gets one "done when" line → Haiku 5.5 = the crew. one task per run, many at once. builds exactly the task, nothing extra → Jev = the inspector. can't write a word. checks every finished task against its "done when" and answers pass, fix or reject → Me = the final yes the rule that made it work: nothing gets merged without a pass from jev and the part people get backwards opus does the least work on the team. it writes the plan, then only sees what jev rejected. it never reads junk honestly the "done when" line did more for quality than any model upgrade. a vague task can't be checked, so it can't be trusted here's the setup. paste it into claude code: build this project as a 3-agent chain. brief: brief.md 1. architect (model: opus): read brief.md once, write plan.md. split the work into tasks of 30 min or less. each tasks/<id>.md gets a goal, inputs and one "done when" line a machine can check 2. builder (model: haiku): one run per task. build only what the task file says. no extra features 3. inspector (jev via typesafe-sdk): for every finished task answer pass / fix / reject against its "done when" line. fix goes back to haiku once. reject goes to opus 4. only "pass" gets merged. log every verdict to verdicts.json 5. show me plan.md first. nothing runs until i say go
This post quotes another post: x.com/xinc3s/status/2108204467774509059