Self-Hosting & Automation
Idea-Council (`/council`) AI Review Board
Parallel multi-seat critique → one senior-dev verdict
Built & Deployed2026
5 read-only review seats in parallel
Merged verdict: risks + stack + MVP + next steps
Foolproof, un-mis-invokable spec
Big decisions benefit from multiple perspectives. This project builds an AI review board: a /council command that fans an idea out to five read-only, role-defined agents in parallel — Critic, Architect, Security, Ops/SysAdmin, and Domain — each told to criticise first and recommend second, then merges their outputs into one senior-dev verdict (risks ranked by severity, recommended stack, MVP plan, ordered next steps, blind spots).
Why it works
- Reduces single-opinion bias — adversarial critique from multiple lenses, not cheerleading.
- Deterministic — a foolproof spec fixes the process so it can’t be invoked ambiguously.
- Read-only & safe — reviewers never implement; they only assess.
The same orchestration that reviewed this portfolio’s architecture is itself a portfolio project — engineering that builds better engineering.
Outcomes
- Built a reproducible AI review board that reduces single-opinion bias for important decisions.
- Encodes a foolproof contract so the process is deterministic and reviewable.
- Used in this very project to critique the architecture before building.
Stack & tools
OpenCode agentsSub-agent orchestrationMCP / toolingPrompt engineeringMarkdown specs & commands