# Prism

> Parallel Review by Independent Specialist Models. Multi-agent review protocol that deploys 5+ specialist reviewers in parallel to eliminate confirmation bias and groupthink.

- Skill: `cogpros/prism` (Agent Skill)
- Install (CLI): `npx skillmds@latest add cogpros/prism`
- Raw SKILL.md: https://api.skillmd.com/api/skills/cogpros/prism/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: cogpros (https://skillmd.com/u/cogpros)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/cogpros/prism

---


# PRISM v2 -- Parallel Review by Independent Specialist Models

PRISM uses parallel independent reviewers to surface issues that consensus-based reviews miss. Disagreements are MORE valuable than consensus.

## When to Use

- Architecture decisions
- Security-sensitive changes
- Major refactors (>500 lines)
- Decisions lasting 6+ months
- Open source releases

**Skip for:** Minor bug fixes, typos, cosmetic changes, urgent hotfixes, easily reversible decisions.

## Three Invocation Modes

| Mode | Agents | Command |
|------|--------|---------|
| **Budget** | 3 (Security, Performance, DA) | "Budget PRISM" |
| **Standard** | 6 specialists | "Run PRISM" |
| **Extended** | 8+ (Standard + Code Reviewers) | "Full PRISM audit" |

## Six Standard Reviewers

1. **Security Auditor** -- Attack vectors, trust boundaries
2. **Performance Analyst** -- Metrics, benchmarks, overhead
3. **Simplicity Advocate** -- Complexity reduction
4. **Integration Engineer** -- Compatibility, migration
5. **Blast Radius Reviewer** -- Downstream system effects
6. **Devil's Advocate** -- Assumptions, risks, regrets (reviews independently, without prior findings)

## How to Execute

### Script-based (portable, any agent):
```bash
~/.openclaw/scripts/prism.sh --artifact <path> [--mode standard|budget|extended]
```

Prompts directory: `~/.openclaw/workspace/.claude/skills/prism/prompts/` (12 templates).

### In-session (Claude CLI):
Launch 5+ parallel Agent subagents, each with a reviewer persona. Synthesize findings. Archive to `analysis/prism/archive/`.

## Critical Rules

- Every finding must cite specific files, line numbers, or command output
- Devil's Advocate receives no prior findings brief (independence by design)
- Assertions without citations are deprioritized
- Disagreements between reviewers often contain the most valuable insights

## Dependencies

- Script engine: `~/.openclaw/scripts/prism.sh`
- Shared lib: `~/.openclaw/scripts/lib/review-common.sh`
- Prompt templates: `~/.openclaw/workspace/.claude/skills/prism/prompts/` (12 files)
- Post-review: `~/.openclaw/scripts/sentry-check.sh`

## Error Recovery

- If a reviewer agent fails, synthesis proceeds with remaining reviewers (minimum 3 for valid review)
- If prompts directory is missing, verify `~/.openclaw/workspace/.claude/skills/prism/prompts/` exists
- If prism.sh exits non-zero, check review-common.sh is sourced correctly