Prism

Multi-perspective agent team analysis plugin for Claude Code.

Prism spawns a coordinated team of specialized AI agents — each analyzing from a different perspective — then cross-validates findings through a Devil's Advocate before producing a final report.

Philosophy

How has humanity solved its hardest problems? Diverse minds in a room, arguing until only the defensible ideas survive. Prism is that room — but the minds are AI specialists, and nobody gets to leave until the weak ideas are dead.

Skills

Skill	Command	Description
analyze	/prism:analyze	General-purpose multi-perspective analysis with MCP-based Socratic verification + ambiguity scoring
incident	/prism:incident	Incident postmortem with 3-6 perspective agents + Devil's Advocate + optional Tribunal
prd	/prism:prd	PRD policy conflict analysis against your reference docs
plan	/prism:plan	Multi-perspective planning with committee debate + consensus enforcement

Prerequisites

Before installing Prism, make sure you have:

1. Claude Code installed and working
2. oh-my-claudecode plugin installed (Prism uses its agent types for team members)

Installation

Step 1: Install the plugin

Inside Claude Code, register the Prism marketplace and install the plugin:

/plugin marketplace add valentin1235/prism
/plugin install prism@prism

Or from the terminal CLI:

claude plugin marketplace add valentin1235/prism
claude plugin install prism@prism

The plugin will be automatically enabled after installation. You can verify with /plugin (Installed tab) or:

claude plugin list

Step 2: Enable Agent Team Mode

Prism uses multi-agent team features (TeamCreate, TaskList, SendMessage, etc.) which require Agent Team Mode to be enabled.

Open ~/.claude/settings.json and add CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS to the env section:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

If you already have an env section with other keys, just add the new key inside it:

{
  "env": {
    "EXISTING_KEY": "existing_value",
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Restart Claude Code after making this change.

Without this setting, Prism skills will refuse to run and show a setup guide instead.

Step 3: Install oh-my-claudecode (agent pack)

Prism does not have its own built-in agents. It currently uses oh-my-claudecode as a general-purpose agent pack, which provides the specialized agent types needed for team analysis (architect, architect-medium, analyst, etc.). The critic role has been replaced by Prism's built-in devils-advocate agent. Install oh-my-claudecode if you haven't already:

/plugin marketplace add Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode@omc

Or from the terminal CLI:

claude plugin marketplace add Yeachan-Heo/oh-my-claudecode
claude plugin install oh-my-claudecode@omc

Step 4: Run setup

Prism includes a setup skill that automatically downloads the prism binary and registers it as a user-scope MCP server:

/prism:setup

This configures the MCP tools (prism_interview) used by the analyze skill's Socratic verification.

Step 5: Verify installation

Restart Claude Code, then type:

/prism:incident

If everything is configured correctly, the skill will start the incident intake process. If Agent Team Mode is not enabled, it will show you the setup instructions.

Full settings.json Example

After completing all installation steps, your ~/.claude/settings.json should contain at least these entries. The env section must be added manually (Step 2), while enabledPlugins are added automatically by claude plugin install:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  },
  "enabledPlugins": {
    "oh-my-claudecode@omc": true,
    "prism@prism": true
  }
}

Usage

Skill	Command	Workflow	Output
analyze	/prism:analyze	Intake → Seed Analysis → Perspective Generation → Parallel Analysts → Socratic Verification (per analyst) → Synthesis	Analysis report
incident	/prism:incident	Intake → Seed Analysis → 3-6 Perspective Agents → Devil's Advocate → optional Tribunal → Report	Postmortem report
prd	/prism:prd path/to/prd.md	Read PRD → 3-6 Policy Analysts → Devil's Advocate → Report	prd-policy-review-report.md
plan	/prism:plan path/to/prd.md	Input Analysis → 3-6 Analysts → Devil's Advocate → Committee Debate (UX + Eng + Planner) → Consensus Loop → Plan	plan.md

All skills share the same core pattern: spawn multi-perspective agents → cross-validate → synthesize. See How It Works for detailed flow diagrams.

analyze: Socratic Verification & Ambiguity Scoring

The analyze skill adds MCP-based Socratic verification on top of the shared pattern. Each analyst self-verifies through a prism_interview loop before reporting:

Analyst → write findings.json → prism_interview(start) → question
       → answer (with evidence) → auto-score → next question
       → ... repeat until PASS or max rounds

Ambiguity Score:

Dimension	Weight	Measures
Evidence Clarity	0.4	Findings backed by concrete evidence (file, line, log)?
Causal Chain Clarity	0.35	Cause-effect chain logically sound and complete?
Recommendation Clarity	0.25	Recommendations specific and actionable?

Verdict: PASS (score meets threshold) or FORCE PASS (max rounds reached, flagged for user attention).

---

Inspired By

Prism's multi-perspective analysis and Socratic verification approach was inspired by Ouroboros.

---

How It Works

Analyze with Socratic Verification (/prism:analyze)

graph TD
    A[User Input] --> B{Prerequisite Gate}
    B -->|Not enabled| X[Show setup guide & STOP]
    B -->|Enabled| C[Problem Intake]

    C --> D["Seed Analysis (Opus)"]
    D -->|"severity, dimensions, research"| E["Perspective Generation (Opus)"]
    E -->|"2-5 lenses from 14 archetypes"| F{User Approval}

    F -->|Modify| E
    F -->|Approve| G[Ontology Scope Mapping]
    G --> H[Spawn Analysts in Parallel]

    H --> I1[Analyst 1: Investigate]
    H --> I2[Analyst 2: Investigate]
    H --> IN[Analyst N: Investigate]

    I1 --> J1["prism_interview Loop"]
    I2 --> J2["prism_interview Loop"]
    IN --> JN["prism_interview Loop"]

    J1 -->|"score ≥ threshold → PASS"| K[Collect Verified Findings]
    J2 -->|"score ≥ threshold → PASS"| K
    JN -->|"score ≥ threshold → PASS"| K

    K --> L[Synthesis & Report]
    L --> M{Complete?}
    M -->|Yes| N[Cleanup]
    M -->|"Deeper Investigation (max 2)"| H

    style D fill:#7c3aed,color:#fff
    style E fill:#7c3aed,color:#fff
    style I1 fill:#4a9eff,color:#fff
    style I2 fill:#4a9eff,color:#fff
    style IN fill:#4a9eff,color:#fff
    style J1 fill:#ef4444,color:#fff
    style J2 fill:#ef4444,color:#fff
    style JN fill:#ef4444,color:#fff
    style F fill:#f59e0b,color:#fff
    style X fill:#dc2626,color:#fff
    style L fill:#16a34a,color:#fff

Loading

Incident Postmortem (/prism:incident)

graph TD
    A[User Input] --> B{Prerequisite Gate}
    B -->|Not enabled| X[Show setup guide & STOP]
    B -->|Enabled| C[Problem Intake]
    C -->|Severity, Evidence, Context| D{SEV1 or Active?}

    D -->|Yes| E1[Fast Track: 4 Core + DA]
    D -->|No| E2[Perspective Generation]
    E2 -->|Select 3-5 archetypes| E1

    E1 --> F[Team Formation]

    F --> G1[Timeline Analyst]
    F --> G2[Root Cause Analyst]
    F --> G3[Systems Analyst]
    F --> GN[+ Extended Analysts]

    G1 --> H[Devil's Advocate]
    G2 --> H
    G3 --> H
    GN --> H

    H --> I{Tribunal needed?}
    I -->|Yes| J1[UX Critic]
    I -->|Yes| J2[Engineering Critic]
    I -->|No| K

    J1 --> K[Final Report]
    J2 --> K

    K --> L[Team Teardown]

    style G1 fill:#4a9eff,color:#fff
    style G2 fill:#4a9eff,color:#fff
    style G3 fill:#4a9eff,color:#fff
    style GN fill:#4a9eff,color:#fff
    style H fill:#ef4444,color:#fff
    style J1 fill:#f59e0b,color:#fff
    style J2 fill:#f59e0b,color:#fff
    style X fill:#dc2626,color:#fff

Loading

PRD Policy Analysis (/prism:prd)

graph TD
    A["/prism:prd path/to/prd.md"] --> B{Prerequisite Gate}
    B -->|Not enabled| X[Show setup guide & STOP]
    B -->|Enabled| C[Read PRD & Sibling Files]
    C --> D[Generate 3-6 Policy Perspectives]

    D --> E[Team Formation]

    E --> F1[Policy Analyst 1]
    E --> F2[Policy Analyst 2]
    E --> F3[Policy Analyst 3]
    E --> FN[Policy Analyst N]

    F1 --> G[All Analysts Complete]
    F2 --> G
    F3 --> G
    FN --> G

    G --> H[Devil's Advocate]
    H -->|Merge, Calibrate, Rank| I["Final Report (prd-policy-review-report.md)"]

    I --> J[Team Teardown]

    style F1 fill:#4a9eff,color:#fff
    style F2 fill:#4a9eff,color:#fff
    style F3 fill:#4a9eff,color:#fff
    style FN fill:#4a9eff,color:#fff
    style H fill:#ef4444,color:#fff
    style X fill:#dc2626,color:#fff

Loading

Plan with Committee Debate (/prism:plan)

graph TD
    A[Input Analysis & Context] --> B[Generate 3-6 Perspectives]
    B --> C[Team Formation]

    C --> D1[Analyst 1]
    C --> D2[Analyst 2]
    C --> D3[Analyst 3]
    C --> DN[Analyst N]

    D1 --> E[Devil's Advocate Synthesis]
    D2 --> E
    D3 --> E
    DN --> E

    E --> F["Committee Debate (UX + Eng + Planner)"]
    F --> G{Consensus?}

    G -->|"3/3 Unanimous"| H["Write plan.md & Cleanup"]
    G -->|"2/3 Working (Normal only)"| H
    G -->|"No Consensus"| I[Shutdown Committee]
    I --> J[Add Perspective + Gap Analysis]

    J -->|"Normal: max 2 loops"| C
    J -->|"Hell: unlimited loops"| C

    style D1 fill:#4a9eff,color:#fff
    style D2 fill:#4a9eff,color:#fff
    style D3 fill:#4a9eff,color:#fff
    style DN fill:#4a9eff,color:#fff
    style E fill:#ef4444,color:#fff
    style F fill:#f59e0b,color:#fff
    style G fill:#d97706,color:#fff,stroke:#333
    style H fill:#16a34a,color:#fff,stroke:#333
    style I fill:#dc2626,color:#fff,stroke:#333

Loading

plan: Consensus Modes

Consensus Level	Condition	Normal Mode	Hell Mode (--hell)
Strong	3/3 agree	Output	Output
Working	2/3 agree	Output	Feedback Loop
No Consensus	<60%	Feedback Loop	Feedback Loop

Normal Mode: max 2 feedback loops, then forced output. Hell Mode: unlimited loops until unanimous.

Agent Mapping

Role	Agent Type	Model
Analyst (complex)	oh-my-claudecode:analyst	opus
Analyst (standard)	oh-my-claudecode:architect-medium	sonnet
Devil's Advocate	prism:devils-advocate	opus
UX / Engineering Critic	oh-my-claudecode:architect / architect-medium	opus / sonnet
Planner	oh-my-claudecode:planner	opus

Troubleshooting

"Agent Team Mode is not enabled"

Add "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" to the env section of ~/.claude/settings.json and restart Claude Code. See Step 2.

"Reference docs not configured"

The skill tried to access reference docs but no brownfield repositories are configured. Run /prism:setup to configure them.

Agents not spawning / TeamCreate fails

Make sure oh-my-claudecode plugin is installed and enabled. Prism's analyst agents depend on oh-my-claudecode agent types (the critic role uses Prism's built-in devils-advocate agent instead). See Step 3.

Skill not showing in autocomplete

Make sure "prism@prism": true is in your enabledPlugins and restart Claude Code.

License

MIT