evaluate

Evaluate Content

A scoring model is a standard of judgment, learned from examples, that can score any content. Each model has traits — distinct dimensions of quality — and scores content 0–100 per trait. Scores are deterministic and instant.

Score content against specific traits. Measurement only — no edits.

Core Flow

1. Identify content to score. Read from a file path ($ARGUMENTS), a URL, a selection, or conversation context. URLs (http/https) can be passed directly to score — they are fetched and text-extracted automatically. If unclear, ask the user what content to evaluate.

2. Determine the model. If the user specifies a model, use it. Otherwise, call list_models to see what's available and either pick the most relevant one based on context (e.g. a README → u22a8.compelling-readme) or present the options and let the user choose.

3. Identify traits. Use list_traits to show the model's traits. Suggest relevant ones based on context, but confirm the user's choice. Do NOT silently evaluate all traits unless the user explicitly asks for "all".

4. Score the content. Call score with the model handle, content text, and specified traits.

5. Present results clearly. For each trait, show:

   • Trait name and score (0–100)
   • The zone label (strong_positive, positive, overlap, negative, strong_negative) — this is the primary interpretation
   • Confidence level (high, moderate, low) — indicates how reliable the score is
   • Headroom — how far from the "solid" threshold (0 means already there)
   • Use the trait's positive/negative labels for human-readable framing (e.g. "Clear" vs "Unclear")

6. Report composite. Show the composite score. If headroom > 0, mention which trait is the bottleneck.

Score Interpretation

Use the zone and breaks from the detail field — these are data-driven from the model, not fixed ranges. The breaks define three thresholds (developing, solid, strong) that separate four labels:

• Strong: score above breaks.strong
• Solid: score between breaks.solid and breaks.strong
• Developing: score between breaks.developing and breaks.solid
• Weak: score below breaks.developing

Present these labels rather than raw numbers when interpreting results.

Important Rules

• Trait selection is explicit. Always confirm which traits the user wants evaluated. Don't default to all.
• Never expose internals. Only report scores and their interpretation — never discuss how scoring works.
• Handle long content. For files, read the full content and pass it as a single text to score. The scoring handles any necessary processing internally.
• Error handling. If a trait name doesn't exist, the MCP tool returns a clear error. Relay it to the user with the list of available traits.
• Use the API's interpretation. Don't invent your own score ranges — use the zones, breaks, confidence, and headroom provided in the detail field.

Example Interaction

User: "Evaluate my README"

1. The content is a README, so u22a8.compelling-readme is the natural fit — use it directly
2. Read the file, call list_traits to show available traits, confirm with user
3. Call score with chosen traits
4. Present using the returned detail:

   Clarity: 72 — Solid (confidence: high)
   Flow: 45 — Developing (confidence: moderate, headroom: 12)
   
   Composite: 56
   Bottleneck: Flow (12 points from solid)
   

5. Add brief qualitative commentary based on the trait's positive/negative labels