Bootstrap Learning

Exit criteria: Can describe the domain in one sentence. Can list 5-10 key terms.

Phase 2: Foundation — Nail the core concepts

Goal: Understand the 3-5 ideas everything else builds on.

• Ask for the simplest possible explanation of each core concept
• Demand concrete examples, not abstractions
• Test understanding by explaining it back in your own words
• Red flag: If the explanation uses jargon from the same domain, you haven't bottomed out

Exit criteria: Can explain core concepts without jargon. Can answer "why does this exist?"

Phase 3: Elaboration — Add depth through cases

Goal: Move from "I understand the concept" to "I can apply it."

Exit criteria: Can identify when to use and when NOT to use the thing.

Phase 4: Connection — Link to existing knowledge

Goal: Integrate new knowledge with what you already know.

• Map analogies: "This is like [existing concept] because..."
• Find contradictions: "This conflicts with [existing belief] — which is right?"
• Identify synergies: "Combining this with [skill X] could improve..."
• Update synapses: Create connections in synapses.json

Exit criteria: At least 2 connections to existing skills identified.

Phase 5: Consolidation — Create persistent memory

Goal: Store the learning in the right format and location.

Exit criteria: At least one memory file created. Synapses updated.

Gap Identification Patterns

Questioning Strategies

Progressive Depth

1. What — "What is X?" (definition)
2. Why — "Why does X exist?" (motivation)
3. How — "How does X work?" (mechanism)
4. When — "When should I use X?" (context)
5. When not — "When should I NOT use X?" (boundaries)

The Feynman Check

If you can't explain it simply, you don't understand it well enough.

After learning a concept, try to explain it in one paragraph using no jargon. If you can't, identify which part is unclear and loop back.

Skill File Quality Bar

A good bootstrap learning output (SKILL.md) should:

• Contain domain knowledge an LLM wouldn't know generically
• Include concrete examples, not just category labels
• Have tables with real data (thresholds, trade-offs, decision criteria)
• Avoid the "capabilities list" anti-pattern ("Expert in X. Can do Y.")
• Pass the Feynman check — any section should be explainable simply

Synapses

See synapses.json for connections.