Multi Ai Review

Multi-AI Review — Universal Consensus Engine

> Heavy-Hitter: State the domain in natural language — right panel assembled automatically. > Cost: CLI subscription plans only — no additional API charges.

Quick Start

# Code review
"Review this PR"

# Market analysis
"How does the market look today?"

# Investment
"Evaluate this investment"

# Consensus (auto-applied)
"Convene a council on this direction"

Prerequisites

command -v claude   # Claude Code (Chairman)
command -v gemini   # Gemini CLI (optional)
command -v codex    # Codex CLI (optional)

---

3-Stage Pipeline

Stage 1: Initial Opinions (parallel)
├── Gemini CLI → opinion.md (perspective A)
└── Codex CLI  → opinion.md (perspective B)

Stage 2: Cross-Review (rebuttal)
├── Gemini reviews/rebuts Codex
└── Codex reviews/rebuts Gemini

Stage 3: Chairman Synthesis
└── Claude synthesizes → Score Card → decides if more rounds needed

---

Domain Auto-Routing

Keywords detected automatically — no preset naming needed.

| Preset | Keywords | Gemini | Codex | |--------|----------|--------|-------| | code-review | review, PR, code, merge | architecture/UX | technical/patterns | | market-regime | market, stocks, macro | macro/news | quant/indicators | | investment | investment, valuation, deal | market/strategy | finance/risk | | risk-assessment | risk, security, danger | external threats | internal weaknesses | | project-gate | gate, milestone, Go/No-Go | stakeholders/scope | schedule/resources | | planning | plan, PRD, design, architecture, council | CTO (architecture/scalability) | Security (threats/compliance) | | default | (fallback) | strategy/opportunity | execution/risk |

Full presets: presets/*.yaml

---

Score Card

Grade Thresholds

| Grade | Min Score | |-------|-----------| | A | 90 | | B | 80 | | C | 70 | | D | 60 | | F | 0 |

Severity Labels

| Label | Meaning | |-------|---------| | Critical | Immediate action | | High | Priority recommended | | Medium | Improvement recommended | | Low | For reference |

Key Dimensions (code-review example)

| Dimension | Weight | |-----------|--------| | Security | 25% | | Performance | 20% | | Maintainability | 25% | | Correctness | 20% | | Style | 10% |

---

Chairman Rules (v4.1.0)

Core Principles:

1. Evidence Hierarchy — file:line citations outrank structural scores
2. Verification Required — score changes need grep/file read/test proof
3. Pre-Deploy Gate — Done-When checks BEFORE deployment
4. Delta Arbitration — score gap ≥15 triggers verification (no averaging)
5. 2× Codex Weight — in code-review/project-gate when verified
6. Council Skeptic — planning domain: play Devil's Advocate. Challenge every recommendation with "What's the tradeoff?" — never accept "good point" without a follow-up probing question.
7. Planning Fallback — planning domain + CLI unavailable: spawn Agent(CTO) + Agent(UX) + Agent(Security) per references/council-personas.md instead of skipping.

Quick Reference:

| Rule | Trigger | Action | |------|---------|--------| | Evidence Hierarchy | Concrete vs structural | Prioritize file:line | | Verification Required | Score adjustment | Run grep/file read first | | Pre-Deploy Gate | Task completion | Block if checks fail | | Delta Arbitration | Gap ≥15 | Verify, don't average | | 2× Weight | code-review + verified Codex | Formula: (codex×2 + gemini)/3 | | Council Skeptic | planning domain | "What's the tradeoff?" on every recommendation | | Planning Fallback | planning + no CLI | Spawn Agent(CTO) + Agent(UX) + Agent(Security) |

Detailed rules: references/evidence-rules.md

---

Chairman Protocol (Stage 3)

1. Extract Evidence
   - File:line citations (Codex)
   - Structural observations (Gemini)
   - Run Done-When verification

2. Consensus Evaluation
   - Consensus reached? → Proceed
   - Unresolved issues? → Additional round
   - Score gap ≥15? → Verify per Evidence Rules

3. Additional Round (if needed, max 2)
   - Focused question on contention point
   - Re-run Cross-Review

4. Final Synthesis
   - Score Card with grades
   - Severity summary
   - Recommendations

Full protocol: references/chairman-protocol.md

---

CI Quality Gate

./skills/multi-ai-review/scripts/council.sh --mode ci "review request"

Thresholds (council.config.yaml):

• Critical issues: fail if > 0
• High issues: fail if > 3
• Overall score: fail if < 70

on_fail: "block" — returns non-zero exit code.

---

Preset Verdicts

| Preset | Verdict Options | |--------|----------------| | code-review | A / B / C / D / F (score-based) | | market-regime | Strong Bull / Bull / Neutral / Bear / Strong Bear | | investment | Strong Buy / Buy / Hold / Pass / Strong Pass | | risk-assessment | Critical / High / Medium / Low / Negligible | | planning | Strongly Recommend / Recommend / Revise / Major Revise / Reject | | default | Strong Agree / Agree / Neutral / Disagree / Strong Disagree |

---

File Structure

skills/multi-ai-review/
├── SKILL.md                    # this file
├── council.config.yaml         # member, routing, scoring config
├── presets/
│   ├── code-review.yaml
│   ├── market-regime.yaml
│   ├── investment.yaml
│   └── default.yaml
├── scripts/
│   └── council.sh              # main execution
└── references/
    ├── evidence-rules.md       # Chairman Evidence Weighting (detailed)
    └── chairman-protocol.md    # Stage 3 protocol (detailed)

---

Last Updated: 2026-04-01 (v4.2.0 — Council mode: planning preset + CTO/UX/Security personas + Council Skeptic rule)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: insightflo
• Source: insightflo/claude-impl-tools
• License: MIT

Install and usage instructions live in the source repository linked above.