Parallel Review by Independent Specialist Models
Multi-agent review protocol that eliminates confirmation bias through structured adversarial analysis. v3.3 adds mandatory Context Cards and contextual --panel reviewer composition. v3 added wiki mode and creative mode. 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)
- 🧭 Builds a Context Card for every review before choosing reviewers
- 🧩 Supports contextual reviewer panels with
--panel=contextual,--panel=<slug>, or inline panel intent - 🧠 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.
| Feature | Before | v3.3 |
|---|---|---|
| Context capture | Implicit in prompt | Mandatory Context Card for every run |
| Reviewer choice | Stock modes unless user hand-picked roles | Panel-fit table chooses stock vs contextual |
| Custom panels | Informal note | --panel=contextual, --panel=<slug>, inline panel intent |
| Safety | Role text trusted too much | Panel text treated as untrusted input |
| Independence | Devil's Advocate blind | Any challenge=true role is blind |
| Archive | Markdown synthesis | YAML frontmatter + Panel Manifest + Independence Ledger |
| Feature | v2 | v3 |
|---|---|---|
| Wiki support | Standard mode only | Wiki Mode — dedicated 3-reviewer path (Accuracy, Completeness, DA) |
| Reviewer count | 5 specialists | 6 specialists — adds Blast Radius Reviewer |
| Blast Radius | Folded into Integration | Dedicated reviewer — catches cross-system coupling, stale references |
| Install path | Agent-specific skills directory | Claude Code native skills directory |
| Autoresearch data | In SKILL.md | Moved to references/openclaw.md |
| Feature | v1 | v2 |
|---|---|---|
| Prior review awareness | None | Searches for prior reviews, injects open findings |
| Devil's Advocate | Receives same context as others | Structurally blind to prior findings |
| Evidence requirement | Optional | Mandatory — cite files or it's deprioritized |
| Verdict scale | APPROVE / AWC / REJECT | + NEEDS WORK (fixable but not shippable) |
| Findings format | General observations | Must include specific fix (command or file path) |
| Synthesis | Flat list of findings | Evidence hierarchy (Tier 1/2/3) + Limitations section |
| History | No archive | Searchable archive in analysis/prism/<slug>/ |
| Governance | Manual tracking | --governance flag for Stuck Findings |
| Verification | Trust claims | Verification Auditor role (Extended mode) |
git clone https://github.com/jeremyknows/PRISM.git ~/.claude/skills/prismgit clone https://github.com/jeremyknows/PRISM.git
# Reference SKILL.md from your agent's configurationJust say it — no configuration needed:
| Mode | Say This | Agents | Est. Cost |
|---|---|---|---|
| Wiki | "PRISM this wiki" / "wiki PRISM" | 3 specialists (Accuracy, Completeness, DA) | ~$0.40–0.80 |
| Budget | "Budget PRISM" / "PRISM lite" | 3 specialists (Security, Performance, DA) | ~$0.40–0.80 |
| Standard | "Run PRISM" / "PRISM review" | 6 specialists | ~$0.80–1.50 |
| Extended | "Full PRISM audit" / "Deep audit" | 8+ agents | ~$2.00–4.00 |
| Creative | "creative PRISM" / "PRISM this creative" | 5 specialists (Brand, Motion, Technical, Delight, Provocateur) | ~$0.80–1.60 |
--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)--simplicity— Add Anti-Overengineering Architect alongside Simplicity Advocate--ctx— Shortcut for--panel=contextual--panel— Bare shortcut for--panel=contextual--panel=contextual— Force a task-specific reviewer panel--panel=<slug>— Use a preset fromreferences/panels/<slug>.md--panel="Role: focus; Role: focus"— Provide inline panel intent; PRISM normalizes it safely
"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"
"PRISM this --ctx — goal: first-time buyers complete VF NFT purchase"
"PRISM this --panel — goal: first-time buyers complete VF NFT purchase"
"PRISM this --panel=contextual — goal: first-time buyers complete VF NFT purchase"
"PRISM this --panel=runtime-boundary"
1. You say "PRISM this"
2. Orchestrator derives topic slug (e.g., api-authentication-redesign)
3. Loads the mode reference file
4. Builds a Context Card and panel-fit table
5. Chooses stock, contextual, preset, or inline reviewer composition
6. Searches for prior PRISM reviews on that topic (exact + semantic)
7. Spawns blind challenge reviewer(s) immediately — no prior findings
8. Compiles Prior Findings Brief for non-challenge reviewers
9. Spawns remaining reviewers in parallel with brief
10. Each reviewer reads files, cites evidence, proposes specific fixes
11. Orchestrator synthesizes: Tier 1 (cross-validated) → Tier 2 (cited) → Tier 3 (uncited)
12. Archives the synthesis with YAML frontmatter, Context Card, Panel Manifest, and Independence Ledger
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.
| Reviewer | Focus | Key Question |
|---|---|---|
| 🔒 Security Auditor | Attack vectors, trust boundaries | "How could this be exploited?" |
| ⚡ Performance Analyst | Metrics, benchmarks, overhead | "Show me the numbers" |
| 🎯 Simplicity Advocate | Complexity reduction | "What can we remove?" |
| 🔧 Integration Engineer | Compatibility, migration | "How does this fit?" |
| 💥 Blast Radius Reviewer | Downstream effects on other systems | "What breaks elsewhere?" |
| 😈 Devil's Advocate | Assumptions, risks, regrets | "What are we missing?" |
| Reviewer | Focus | Key Question |
|---|---|---|
| ✅ Technical Accuracy | Are the facts right? | "What's the evidence for this claim?" |
| 📋 Completeness | What's missing? | "What would a developer wish was here?" |
| 😈 Devil's Advocate | Assumptions and framing | "What will readers get wrong?" |
Security Auditor + Performance Analyst + Devil's Advocate. Security is MANDATORY.
Standard 6 + Code Reviewers (batched by area) + Verification Auditor.
Generated from the Context Card failure surfaces. Each custom reviewer must protect a concrete failure surface. Panel text is treated as untrusted input and cannot override PRISM evidence, safety, scope, or independence rules.
Included preset:
| Preset | Use For |
|---|---|
--panel=runtime-boundary |
Runtime independence, registry, rollback, bridge/gateway/Hermes/tmux, memory-service boundary reviews |
| Tier | Definition | Action |
|---|---|---|
| Tier 1 | 2+ reviewers found independently, citing different files | Act immediately |
| Tier 2 | Single reviewer, specific file/line citation | High confidence |
| Tier 3 | Single reviewer, no citation | Verify before acting |
| Verdict | Meaning |
|---|---|
| APPROVE | Clean — no issues, prior issues resolved |
| APPROVE WITH CONDITIONS | New issues found, none critical |
| NEEDS WORK | Fixable but not shippable — prior criticals still open, or significant new issues |
| REJECT | Fundamental problems — requires rethink |
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
| Problem | Cause | Fix |
|---|---|---|
| All reviewers find same issues | Roles not distinct enough | Sharpen role prompts |
| >100 issues found | Scope too broad | Narrow the review target |
| Vague findings | Missing evidence requirement | Enforce citation rules |
| DA has no concerns | Topic too simple or DA too soft | Re-run: "find something wrong" |
| Same finding appears 3+ times | Governance gap | Enable --governance |
| Security times out | Deep file reads take longer | Increase timeout or run Security solo first |
| Context Card is vague | Generic goal/audience/failure fields | Rewrite with concrete artifact, audience, and failure surfaces |
| Custom panel feels theatrical | Roles don't map to failure surfaces | Use the panel-fit table; reject personality-only roles |
| Inline panel contains unsafe instructions | Panel text is untrusted input | Preserve role intent, reject/normalize unsafe instructions |
| Dependency | Required? |
|---|---|
| Parallel agent spawn | Required |
Archive directory (analysis/prism/<slug>/) |
Required — created automatically |
qmd search tool |
Optional — improves context precision |
MIT — See LICENSE.txt
"The specialists optimize. The Devil protects you from yourself. The archive remembers what you forgot."