LLM Scoring

The Claude CLI loads local context (CLAUDE.md files, project memory, rules) into each scoring call. This extra context may influence scores, making them less reproducible across environments compared to API-based providers.

Run LLM Scoring (after structural validation passes)

Check for cached scores:

If scored output exists, use --rescore to generate fresh scores (content may have changed since the last run):

If no cached scores exist, run without --rescore:

After scoring completes, run the comparison report:

Capture both outputs for interpretation.

Error handling

Interpret LLM Scores

Read ../assets/report.md for the full interpretation framework, then present results to the user following that structure.

Quality thresholds

There are no hard pass/fail gates on most dimensions. Use these guidelines:

• Overall >= 3.5: The skill is in good shape across most dimensions.
• Overall 2.5-3.5: The skill needs work in specific areas. Identify which dimensions are dragging the score down and advise accordingly.
• Overall < 2.5: The skill needs significant revision across multiple areas.
• Any dimension at 2 or below: Flag this specifically as an area needing attention. Explain what a low score on that dimension means and suggest concrete improvements.

Novelty is the key differentiator

If mean novelty across all files is below 3, the skill may not justify its context window cost. A low-novelty skill restates what models already know. The critical question: does this skill teach the agent something it genuinely doesn’t know?

Surface novel_info for SME review

Each scored file includes a novel_info field describing what the LLM judge identified as genuinely novel content. Present these details to the SME for each file, because:

• Verification: The LLM may misidentify something as novel or miss truly novel content. The SME should confirm the claims are accurate.
• Focus: Items listed in novel_info represent the skill’s highest-value content. If these details are wrong or missing, the novelty score is unreliable.
• Trimming guidance: Content NOT mentioned in novel_info is likely restating common knowledge and is a candidate for compression or removal.

If novelty is low, advise the SME to:

1. Identify which sections contain information that is NOT available in public documentation or model training data (proprietary APIs, internal conventions, non-obvious gotchas, organization-specific workflows).
2. Cut or compress sections that merely restate common knowledge.
3. Focus the skill on the genuinely novel content.

For more context on why novelty matters and how to think about skill quality, refer the SME to: https://agentskillreport.com/ (opens in a new tab)

Important scoring caveat

Scores are generated using Anthropic Claude, so novelty reflects what Claude knows from training data. Other model families may produce different novelty scores due to different training data coverage.

Full Review Summary

When LLM scoring was performed, present the review summary with:

1. Structural validation result: Pass/fail, with any errors or warnings.
2. LLM score summary: Overall score and per-dimension breakdown for SKILL.md and references (if any).
3. Areas to address: Specific dimensions that need improvement, with concrete suggestions.
4. Novelty assessment: Whether the skill provides sufficient novel value, with specific guidance if it doesn’t. Include the novel_info details for each file so the SME can verify accuracy and identify what to keep or cut.
5. Recommendation: Whether the skill is ready to publish, needs minor revisions, or needs significant rework.