gh-tommy-ca-notion-skills-plugins-notion-skills/skills/knowledge-capture/evaluations/README.md

Knowledge Capture - Evaluation Scenarios

This directory contains test scenarios for validating the Knowledge Capture skill across different Claude models.

Overview

These evaluations ensure that Knowledge Capture works consistently and effectively with:

• Claude 3 Haiku
• Claude 3 Sonnet
• Claude 3 Opus

Each scenario includes:

• Input: The conversation or content to capture
• Expected Behavior: What the skill should do
• Success Criteria: How to verify success

Running Evaluations

Manual Testing

1. Copy the conversation/input from a scenario
2. Activate Knowledge Capture skill in Claude Code
3. Request to capture it as specified
4. Evaluate against the success criteria

Test Coverage

• Basic conversation capture
• Different documentation types
• Complex multi-topic discussions
• Action item extraction
• Decision documentation
• Error handling

Scenarios

scenario-1-meeting-notes.json

Type: Meeting Summary Capture Complexity: Medium Input: Unstructured meeting notes from a team discussion Expected Output: Structured meeting summary with decisions and action items Success Criteria: All action items identified, decisions documented, participants listed

scenario-2-architecture-discussion.json

Type: Decision Record Capture Complexity: Medium-High Input: Complex technical architecture discussion Expected Output: Decision record with options considered and rationale Success Criteria: All alternatives documented, clear decision stated, consequences identified

scenario-3-quick-context.json

Type: Quick Capture Complexity: Low Input: Brief status update Expected Output: Quick reference document Success Criteria: Key points captured, properly formatted

scenario-4-action-items-focus.json

Type: Action Item Extraction Complexity: Medium Input: Meeting with many scattered action items Expected Output: Clear list of who, what, when Success Criteria: All action items found, owners assigned, deadlines noted

scenario-5-learning-from-incident.json

Type: Learning Document Complexity: High Input: Post-incident discussion about what went wrong Expected Output: Learning document with best practices Success Criteria: Root causes identified, lessons extracted, preventive measures noted

Evaluation Criteria

Comprehensiveness

• Did the skill capture all key information?
• Were important details missed?
• Is the output complete?

Structure

• Is the output well-organized?
• Are headings and sections appropriate?
• Does it follow the documentation type's structure?

Actionability

• Are action items clear and assignable?
• Are next steps defined?
• Can someone understand what to do next?

Linking

• Are cross-references included?
• Could this be discovered from related docs?
• Are important connections made?

Accuracy

• Is the captured information accurate?
• Were details misinterpreted?
• Does it faithfully represent the source?

Model-Specific Notes

Haiku

• May need slightly more structured input
• Evaluates on accuracy and completeness
• Good for quick captures

Sonnet

• Handles complex discussions well
• Good at inferring structure
• Balanced performance across all scenario types

Opus

• Excels at inferring structure from unstructured input
• Best for complex, multi-topic discussions
• Highest accuracy on nuanced extraction

Failure Modes to Watch For

1. Orphaning: Documentation created without linking to hub pages
2. Structure Loss: Loss of hierarchical organization
3. Missing Details: Key decisions or action items not captured
4. Wrong Type: Documentation created as wrong type (e.g., FAQ instead of Decision Record)
5. Broken Context: Missing important contextual information
6. Inaccuracy: Misinterpretation of what was discussed

Updating Scenarios

When updating scenarios:

1. Ensure they reflect real-world use cases
2. Include edge cases and challenging inputs
3. Keep success criteria clear and measurable
4. Test with all three model sizes
5. Document any model-specific behavior

Expected Performance

Target Success Rate: 90%+ on Sonnet and Opus, 85%+ on Haiku

When a scenario fails:

1. Review the specific failure mode
2. Check if it's a consistent issue or model-specific
3. Update the skill prompt if needed
4. Document the failure and resolution