10 Runs, 8 Signed Off, 7 Passed the Real Gate: Three Claude Code Seats Argue Over the Spec Before Anyone Writes Code
- Level Advanced
- Setup1–3 hrs (estimated)
- Evidence Some evidence
- CostNot stated in source
- No-codeNo
What did they do with Claude Code?
A developer kept getting code that was well written for the wrong plan, so he built a setup where three AI seats argue over a product brief before anyone writes code: a Lead PM writes the brief, a Chief Engineer checks every claim against the real repo and has to cite a file and line for each objection, and a Chief Architect with tools turned off rules on deadlocks and reviews the agreed document. He approves the signed-off document himself; only then does an Engineer implement it in an isolated clone, and a harness runs the project's real test commands. Last week he ran ten councils in solo mode, meaning every seat was Claude Code (Lead PM on Sonnet, Chief Engineer on Opus, Chief Architect on Sonnet with tools off), with at most 4 rounds, on local copies of a toy slugify library and the tool itself. 8 of 10 signed off. 7 of those 8 passed their gate. Two stalled after the Architect raised a late concern and the revision budget ran out. One signed off and then ended BLOCKED because the project's configured npm test was still broken, even though the plan's own tests passed. The median time to a sign-off or a stall was 6.5 minutes. In all six slugify runs, the Chief Engineer's first objection was that npm test ran 0 of 2 existing tests on Node 22. In a later comment he reported a single-agent baseline on the same 10 tasks: the lone agent also caught and fixed the broken test script in all 6 playground runs, and it passed the real gate 10 of 10 times, against the council's 7 of 10. He says that on small, well-specified tasks the debate did not change outcomes, and that the fair test is harder, vaguer tasks.
How it went
dev.to post by Ashkan Goshtasb (Nov 7, 2026) about Open Council, an MIT-licensed CLI he built. Table of 10 solo-mode Claude Code runs on version 0.2.1: 8 signed off, 7 of those 8 passed the project gate, 2 stalled, 1 BLOCKED; median 6.5 minutes. He later commented that a single-agent baseline on the same 10 tasks passed 10 of 10 and also caught the broken npm test script in all 6 playground runs. He discloses the article was written with AI help and the runs were driven by an AI agent; he checked the numbers against the run logs
Worth knowing
- dev.to post by Ashkan Goshtasb (Nov 7, 2026) about Open Council, an MIT-licensed CLI he built. Table of 10 solo-mode Claude Code runs on version 0.2.1: 8 signed off, 7 of those 8 passed the project gate, 2 stalled, 1 BLOCKED; median 6.5 minutes. He later commented that a single-agent baseline on the same 10 tasks passed 10 of 10 and also caught the broken npm test script in all 6 playground runs. He discloses the article was written with AI help and the runs were driven by an AI agent; he checked the numbers against the run logs.
Prompt
Cleaned version rewritten from the source — not a direct quote.
Before any code is written for [task], run a review with three roles. Lead PM: write a short brief and a PRD. Chief Engineer: check every claim against the real repo; every objection must cite a file and line. Chief Architect: review the document only, with no repo tools, and rule on disagreements. Do not implement anything until I approve the signed-off document. After I approve, implement in an isolated copy and run the project's actual test command: [npm test or the equivalent]. If that gate fails, mark the work BLOCKED, even if your own extra tests passed. If reviewers still disagree when the round limit is hit, stop and save the last draft plus the open objections. Do not call the work done yourself.
Recommended
Try it in Grok Bot
- Copy the prompt (button below).
- Open Grok Bot chat and paste it. Replace any [bracketed] placeholders.
- Connect the listed tools (as needed) as connectors when Grok Bot asks.
- If the task repeats, ask: "Make this a daily routine."
Alternatives: the same prompt works in Claude Code, Claude, ChatGPT Agent, OpenClaw.
Set it up
Stack: Claude Code
- Read the original source for full context.
- Connect the tools your agent needs.
- Paste the prompt and replace [bracketed] placeholders.
- Run once manually and check the output before scheduling it.
Generic suggested steps, not from the source.
Evidence
dev.to post by Ashkan Goshtasb (Nov 7, 2026) about Open Council, an MIT-licensed CLI he built. Table of 10 solo-mode Claude Code runs on version 0.2.1: 8 signed off, 7 of those 8 passed the project gate, 2 stalled, 1 BLOCKED; median 6.5 minutes. He later commented that a single-agent baseline on the same 10 tasks passed 10 of 10 and also caught the broken npm test script in all 6 playground runs. He discloses the article was written with AI help and the runs were driven by an AI agent; he checked the numbers against the run logs (as of 2026-11-07)
Source: Web · 2026-11-07 · Verified by us: not yet · Report broken or outdated