Parallel Review by Independent Specialist Models

Multi-agent review protocol that eliminates confirmation bias through structured adversarial analysis. v3 adds wiki mode and creative mode — a 5-role path for brand and motion direction reviews with Brand Creative Memory — a targeted 3-reviewer path for documentation accuracy. v2 added memory — the system learns from its own review history.

• 🔒 Deploys 6 specialist reviewers in parallel (Security, Performance, Simplicity, Integration, Blast Radius, Devil's Advocate)
• 🧠 Remembers — searches for prior reviews on the same topic, tracks what was fixed
• 📚 Wiki Mode — dedicated 3-reviewer path for documentation accuracy, completeness, and framing
• 🎯 Surfaces disagreements as the most valuable signal
• 😈 Devil's Advocate reviews blind — no prior findings, guaranteed independence
• 📋 Requires evidence — every finding must cite a specific file, line, or command output
• 🔧 Findings include shell commands or file paths — actionable, not just advisory
• ⚡ Works with parallel subagents OR sequential single-agent review

"Disagreements are MORE valuable than consensus."

When 4/5 reviewers agree and 1 dissents, pay attention to that dissent.

"Findings without evidence are noise."

Every finding must cite a specific file. Assertions without citations are lowest priority.

"A finding flagged 3 times tells you about governance, not about the code."

PRISM tracks how many times a finding appears — repeat findings escalate automatically.

git clone https://github.com/jeremyknows/PRISM.git ~/.claude/skills/prism

git clone https://github.com/jeremyknows/PRISM.git
# Reference SKILL.md from your agent's configuration

Just say it — no configuration needed:

• --opus — Use Opus model (critical decisions, ~2-3x cost)
• --haiku — Use Haiku model (fast sanity checks)
• --governance — Surface Stuck Findings (flagged 3+ times without resolution)

"PRISM this wiki article"
"wiki PRISM on satori-og-edge.md"
"PRISM this API change"
"Budget PRISM on the auth flow"
"Full PRISM audit --governance — we've reviewed this area before"
"creative PRISM on the hero animation brief"
"brand review PRISM — veefriends launch video"

1. You say "PRISM this"
2. Orchestrator derives topic slug (e.g., api-authentication-redesign)
3. Searches for prior PRISM reviews on that topic (exact + semantic)
4. Spawns Devil's Advocate immediately (blind — no prior findings)
5. Compiles Prior Findings Brief for remaining reviewers
6. Spawns all other reviewers in parallel with brief
7. Each reviewer reads files, cites evidence, proposes specific fixes
8. Orchestrator synthesizes: Tier 1 (cross-validated) → Tier 2 (cited) → Tier 3 (uncited)
9. Archives the synthesis for future reviews

First run: No prior findings exist. PRISM runs cleanly — no empty sections, no confusion.

Subsequent runs: Prior findings are injected. Reviewers verify what's fixed and hunt for what's new. DA stays blind — their independence is the control group.

Security Auditor + Performance Analyst + Devil's Advocate. Security is MANDATORY.

Standard 6 + Code Reviewers (batched by area) + Verification Auditor.

Don't:

• ❌ Let reviewers see each other's findings (groupthink)
• ❌ Give Devil's Advocate the Prior Findings Brief (breaks independence)
• ❌ Accept findings without file citations (noise)
• ❌ Skip synthesis (raw findings aren't actionable)
• ❌ Skip archiving (breaks memory for future reviews)

Do:

• ✅ Spawn DA immediately, other reviewers after brief is ready
• ✅ Give each reviewer narrow focus (depth > breadth)
• ✅ Require citations in every finding
• ✅ Archive every synthesis to analysis/prism/<slug>/
• ✅ Run two rounds for important decisions

MIT — See LICENSE.txt

"The specialists optimize. The Devil protects you from yourself. The archive remembers what you forgot."