hellblazer/substantive-critique

Deep Critique Skill

Delegates to the substantive-critic agent. See registry.yaml for details.

When This Skill Activates

• After completing a design document or specification
• When reviewing architectural decisions
• After mcp__plugin_nx_nexus__nx_plan_audit for additional depth (RDR-080)
• When validating implementation against specification
• When cross-referencing documentation for consistency
• When verifying claims against evidence
• Before major milestones or releases

Pre-Dispatch: Seed Link Context (optional)

If the critique targets an RDR or architectural artifact, seed link-context so critique findings stored to T3 auto-link. See /nx:catalog for details.

Model Selection

The substantive-critic agent defaults to haiku (fast, cheap). Escalate at dispatch time via the Agent tool's model parameter:

| Task Shape | Model | When | |-----------|-------|------| | Single file, focused critique | haiku (default) | README review, small doc, single function | | Multi-file or cross-RDR | sonnet | >3 files, plan critique, implementation vs spec | | Cycle-level or architectural | opus | Multi-RDR assessment, system-wide design review |

Pass model: "sonnet" or model: "opus" on the Agent tool call to override.

Agent Invocation

Use the Agent tool to invoke substantive-critic:

## Relay: substantive-critic

**Task**: [what needs to be done]
**Bead**: [ID] or 'none'

### Input Artifacts
- Files: [relevant files]

### Deliverable
Critique report with prioritized findings

### Quality Criteria
- [ ] All major dimensions analyzed (structure, logic, completeness)
- [ ] Findings prioritized by impact
- [ ] Recommendations are specific and actionable

For full relay structure and optional fields, see RELAY_TEMPLATE.md.

Critique Methodology

The agent uses mcp__plugin_nx_sequential-thinking__sequentialthinking:

1. Establish context and purpose of artifact
2. Identify criteria/specifications it should conform to
3. Gather evidence from nx store (search tool: corpus="knowledge") and related artifacts
4. Analyze structural integrity
5. Analyze logical consistency
6. Assess completeness
7. Synthesize findings by priority (Critical/Significant/Minor)
8. Formulate actionable recommendations

Success Criteria

• [ ] Context and purpose established
• [ ] Evidence gathered from relevant sources
• [ ] All major dimensions analyzed (structure, logic, completeness)
• [ ] Findings prioritized by impact
• [ ] Recommendations are specific and actionable
• [ ] No surface-level issues dominate over substantive ones
• [ ] T2 memory updated with session findings (if multi-session work)

Agent-Specific PRODUCE

Outputs generated by the substantive-critic agent:

• T3 knowledge: Critique findings via store_put tool: content="# Critique: {artifact}\n{findings}", collection="knowledge", title="critique-{component}-{date}", tags="critique,review"
• T2 memory: Session findings via memory_put tool: content="content", project="{project}", title="critique-{date}.md", ttl="30d"
• T1 scratch: Critique notes during analysis via scratch tool: action="put", content="Critique note: {finding}", tags="critique,review" (promoted to T2 on completion)

Session Scratch (T1): Agent uses scratch tool for ephemeral working notes during the session. Flagged items auto-promote to T2 at session end.