Nw Review Workflow

Agent Review Workflow

V2 Validation Checklist (11 Points)

Run against every agent under review. Each item pass/fail.

• Frontmatter format: --- delimited YAML with name and description
• Size compliance: Under 400 lines; domain knowledge in Skills
• Divergence-only: Only behaviors diverging from Claude defaults
• Calm language: No "CRITICAL", "MANDATORY", "ABSOLUTE"
• Examples present: 3-5 canonical for critical/subtle behaviors
• Least privilege tools: Minimum needed in frontmatter
• maxTurns set: Present in frontmatter
• Platform safety: Via frontmatter/hooks, not prose
• Affirmative phrasing: "Do X" not "Don't do Y"

10. Consistent terminology: One term per concept

11. Clear delegation: Description states when to delegate

Scoring Methodology

Per-Dimension Scoring

For each of 7 critique dimensions (from critique-dimensions skill):

• Pass: All checks satisfied
• Fail: One or more checks not satisfied

Verdict Logic

IF any high-severity dimension fails:
  verdict = "revisions_needed"
ELIF count(medium-severity failures) >= 3:
  verdict = "revisions_needed"
ELSE:
  verdict = "approved"

High-severity: template_compliance, size_and_focus, safety_implementation, priority_validation

Medium-severity: divergence_quality, language_and_tone, examples_quality

Evidence Requirements

Every finding includes: Dimension (which of 7) | Severity (high/medium/low) | Finding (observed, with line numbers/counts) | Recommendation (specific fix action)

Common V1 to V2 Migration Issues

| Residual Pattern | What to Flag |

|-----------------|-------------|

| Embedded YAML config blocks | Should be frontmatter or removed |

| activation-instructions section | Remove -- Claude Code handles activation |

| IDE-FILE-RESOLUTION section | Remove -- not needed in v2 |

| commands with 10+ entries | Reduce to 3-5 focused |

| Inline embed_knowledge | Extract to Skills |

| 5+ "production frameworks" | Remove -- platform handles safety |

| CRITICAL: prefixed instructions | Rephrase as calm direct statements |

| Python/YAML safety code examples | Remove -- aspirational, not executable |

Peer Review Protocol

• Receive agent file path from builder
• Execute full dimension review
• Return structured YAML verdict
• If revisions_needed, include prioritized fix list
• Builder revises and resubmits (max 2 iterations)
• On second rejection, escalate to user

Command Template Review

Additional checks for nWave command files (tasks):

• Size: 50-60 lines target; >60 warning, >150 major, >500 blocker
• Structure: agent activation metadata, context files section, success criteria
• Delegation: business logic belongs in agent, not command
• No embedded procedural steps (STEP 1, STEP 2)