prompt-reviewer

name	prompt-reviewer
description	Internal reviewer agent. Validates a completed prompt folder against metadata.json and models registry. Checks consistency, completeness, and production-readiness. Not user-invocable — triggered automatically by the prompt-creator after convergence completes.
user-invocable	false

Prompt Reviewer

Validate a completed prompt folder for production-readiness. This skill runs automatically after convergence — do not wait for user input.

Input

Read the prompt folder at the path provided. The folder contains:

• prompt.<format> — the prompt file
• metadata.json — scores, model, techniques, config
• tests.json — regression test cases
• report.pdf — generated audit report

Validation Checks

Execute ALL checks. Report results as PASS/FAIL with details.

1. File Completeness

• prompt.<format> exists and is non-empty
• metadata.json exists and is valid JSON
• tests.json exists and has >= 3 test cases
• report.pdf exists and is > 0 bytes

2. Metadata Consistency

Read metadata.json and verify:

• target_model exists in ${CLAUDE_PLUGIN_ROOT}/../../shared/models-registry.json
• scores.overall matches the average of the 5 axis scores
• tokens.estimated is a positive number
• tokens.context_window matches the registry value for the target model
• status is "pass" if overall >= 6, "needs_improvement" otherwise
• version is a positive integer
• config.temperature is appropriate for the domain (0-0.3 for coding, 0.7-1.0 for creative)

3. Prompt-Metadata Alignment

• Prompt format (file extension) matches metadata.format
• If target model prefers XML and format is XML, verify prompt contains XML tags
• If target model prefers Markdown and format is MD, verify prompt contains headers
• Techniques listed in metadata are actually reflected in the prompt structure

4. Score Validation

Run self-eval on the prompt and compare with metadata scores:

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/self-eval.py <prompt-file>

• Scores from self-eval match metadata.scores (tolerance: ±1 per axis)
• If scores diverged significantly, metadata is stale — flag for update

5. Registry Cross-Reference

Read the model entry from ${CLAUDE_PLUGIN_ROOT}/../../shared/models-registry.json:

• If model has reasoning: "reasoning-native", verify no CoT in prompt
• If model has few_shot: "REQUIRED", verify examples exist in prompt
• If model has few_shot: "AVOID", verify no examples in prompt
• If model format is xml, verify prompt uses XML tags
• If model format is markdown, verify prompt uses Markdown headers

Output

Report in this format:

REVIEW: <prompt-name>
  PASS  File completeness (4/4 files)
  PASS  Metadata consistency (7/7 checks)
  FAIL  Prompt-metadata alignment: format is .md but model prefers XML
  PASS  Score validation (within tolerance)
  PASS  Registry cross-reference (4/4 checks)

VERDICT: PASS (4/5 checks) — 1 issue found
ACTION: Convert prompt format from Markdown to XML for Claude target model

If all checks pass: VERDICT: APPROVED — prompt is production-ready If any FAIL: list the specific fixes needed. The convergence engine or main agent should apply them.