learning-capture（学习捕捉）

Learning Capture

Overview

This skill enables continual learning by recognizing valuable patterns during work and capturing them as new skills. It focuses on high-ROI captures: patterns that will save significant context window tokens through frequent reuse.

Recognition Framework

Monitor for these five types of learning moments:

1. Novel Problem-Solving Approaches

Trigger: Develop a creative, non-obvious solution to a complex problem that could apply to similar future problems.

Strong signals:

• Solution required multi-step reasoning or novel tool combinations
• Approach is generalizable beyond this specific instance
• User expresses satisfaction with the results
• Similar problem type likely to recur

2. Repeated Patterns

Trigger: User requests similar tasks 2-3 times and a consistent approach emerges.

Strong signals:

• Pattern has repeated 2+ times with consistent structure
• User asks "can you do the same thing as before?"
• Task type is clearly ongoing (e.g., weekly reports, monthly communications)
• Each instance requires re-explaining the approach

3. Domain-Specific Knowledge

Trigger: User explains company processes, terminology, schemas, or standards that span multiple conversations.

Strong signals:

• Information accumulates across 2+ conversations
• Knowledge is stable (won't change weekly)
• User frequently asks questions in this domain
• Re-explaining costs 1000+ tokens each time

4. Effective Reasoning Patterns

Trigger: Discover a particular way of structuring thinking that consistently produces better results.

Strong signals:

• Pattern applies to a category of problems, not just one instance
• Results are notably better than simpler approaches
• Structure is teachable and reproducible
• Problem category recurs frequently

5. Workflow Optimizations

Trigger: Figure out an efficient way to chain tools or steps together that produces comprehensive results.

Strong signals:

• Workflow chains 3+ distinct steps
• Pattern generalizes to similar task types
• User appreciates the thoroughness
• Similar workflows likely needed regularly

Decision Framework

Offer capture when ALL of the following are true:

1. High confidence (>95%) of significant ROI:

- Pattern will be reused 10+ times across future conversations - Each reuse saves 500+ tokens of re-explanation - The skill itself costs <5000 tokens to load

1. Strong reusability signal present:

- Pattern has repeated 2+ times already, OR - User explicitly indicates ongoing need ("I do this weekly"), OR - Complex domain knowledge worth formalizing, OR - Novel workflow with clear generalizability

1. Not redundant with existing capabilities:

- No existing skill already covers this pattern - Adds meaningful value beyond general knowledge

Do NOT offer capture when:

• First instance of a pattern (wait for repetition)
• Highly context-specific solution (won't generalize)
• Simple task using existing capabilities (no marginal value)
• Creative/one-off work (low reuse probability)
• Ambiguous reusability (unclear if it will recur)

Consult references/decision-examples.md for concrete examples of high-confidence vs. low-confidence scenarios.

Capture Process

Step 1: Recognize the Learning Moment

While working, monitor for recognition triggers from the framework above. Track:

• Is this a repeated pattern?
• Does this generalize beyond this instance?
• Would formalizing this save significant tokens in future uses?

Step 2: Evaluate Against Decision Framework

Before offering capture, verify:

• ROI calculation: (Expected_reuses × Tokens_saved) >> Skill_cost
• Strong reusability signal is present
• Not redundant with existing capabilities

If all checks pass, proceed to offer. If uncertain, do NOT offer.

Step 3: Offer Capture Conservatively

Timing: Offer after completing the immediate task, not mid-task.

Phrasing: Be concise and specific about what would be captured and why it's valuable.

Good examples:

• "I notice I've structured the last three internal comms documents similarly. Would it be helpful to capture this as a skill for future communications?"
• "I've built up understanding of your data architecture across our conversations. Should I formalize this as a skill for more efficient future reference?"
• "The validation workflow I developed seems applicable to your other messy datasets. Worth capturing as a skill?"

Avoid:

• Over-explaining the decision reasoning
• Offering when confidence is <95%
• Interrupting task flow to offer

Step 4: Structure the Draft Skill

When user agrees to capture, create a draft skill file following these steps:

1. Select appropriate template from references/skill-templates.md based on learning moment type
2. Structure the skill using the template as a guide
3. Keep it concise: Focus on what's non-obvious and reusable
4. Include specific triggers: Make it clear when to use this skill
5. Add examples where helpful for clarity
6. Save to outputs: Create the draft at /mnt/user-data/outputs/[skill-name].skill/

The draft skill should be ready for user review and upload with minimal editing needed.

Step 5: Present the Draft

After creating the draft skill:

1. Provide context: Briefly explain what the skill captures and why it will be valuable
2. Highlight key sections: Point out the most important parts of the skill
3. Suggest refinements: Note any areas where user input would improve the skill
4. Explain next steps: User reviews, potentially edits, then uploads via the UI for future conversations

Key Principles

Conservative by default: Better to capture 80% of truly valuable patterns than create noise. Only offer when confidence is very high.

ROI-focused: Prioritize patterns with high reuse frequency and high token savings per reuse.

Context window awareness: Skills cost tokens to load. A skill should pay for itself within 10 uses.

Interpretable: Skills are plain text and easy to review, correct, and refine. This transparency is a feature.

User-controlled: The manual upload step ensures quality control and user agency over what gets added to the knowledge base.

Resources

references/skill-templates.md

Templates for structuring different types of skills based on the learning moment type. Includes:

• Workflow/Process skill template
• Domain Knowledge skill template
• Task Pattern skill template
• Reasoning/Prompt Pattern skill template
• Template selection guide

Read this file when structuring a captured skill to use the appropriate template.

references/decision-examples.md

Detailed examples of high-confidence capture scenarios (where to offer) and low-confidence scenarios (where NOT to offer). Includes:

• Concrete examples with signal analysis
• Recognition pattern checklists
• Decision threshold guidelines
• ROI calculation examples

Read this file when uncertain whether a learning moment meets the capture threshold.