auto-learn
Autonomous web-based self-learning engine. Detects knowledge gaps, searches the web, reads, verifies, synthesizes, and stores learnings. Triggers automatically when confidence is low, evidence is thin
SKILL.mdskills/imported/auto-learn/SKILL.md
GitHubAuto-Learn — Autonomous Self-Learning Engine
"不知道 → 自己查 → 自己学 → 自己记住。不等用户教。"
Core Loop
Detect gap → Search → Read → Verify → Synthesize → Store
↑ │
└──────────── continuous improvement ────────────────┘
Trigger: Knowledge Gap Detection
Auto-learn fires when any of:
| Signal | Meaning | Example |
|---|---|---|
| Confidence LOW | Brain is guessing, not knowing | "I think this API works like..." |
| No evidence | Making a claim without a source | "This library is the fastest" without data |
| Unfamiliar concept | First time encountering a tool/term | User mentions a new framework |
| Version uncertainty | Might be outdated knowledge | "This changed in v3... right?" |
| User challenge | User pushes back on a claim | "Are you sure about that?" |
| Trending topic | Something new in the last 6 months | Recent releases, new standards |
Process
Step 1: Detect & Formulate
Turn the knowledge gap into 2-3 search queries:
Gap: "Not sure about the current state of X"
Query 1: "X latest version 2026"
Query 2: "X vs alternatives comparison 2026"
Query 3: "X common pitfalls 2026"
Step 2: Search & Fetch (parallel)
WebSearch(query1) ─┐
WebSearch(query2) ─┼── parallel ──→ Fetch top 1-2 results each
WebSearch(query3) ─┘
Step 3: Cross-Verify
Don't trust a single source. Check:
- Do 2+ independent sources agree?
- Is the source authoritative? (official docs > blog > forum)
- Is the information recent? (within 6 months for fast-moving tech)
- Any contradictions between sources?
Step 4: Synthesize
Write a 3-5 sentence summary:
Topic: [what was learned]
Key facts: [3-5 bullet points with source references]
Certainty: HIGH / MEDIUM / LOW
Applicable to: [what problems does this knowledge solve?]
Step 5: Store
memory-db.ps1 store
- EntityName: [topic slug]
- EntityType: "learning"
- Content: [synthesized summary + sources]
- Score: 0.6 (learned but not yet applied; will increase when used)
When NOT to Use
| Skip | Reason |
|---|---|
| Standard language syntax | Training data is sufficient |
| Well-known stable libraries | No version uncertainty |
| Opinion/preference questions | Not a knowledge gap |
| Already stored with score ≥0.8 | No need to re-learn |
Integration
- orchestrator: L2+ tasks with knowledge gaps → orchestrator routes to auto-learn before execution
- context7-proactive: Auto-learn finds concept → context7 dives into API specifics
- memory-lora: All learnings stored with sources, cross-referenced
- second-brain: Brain 2 verifies auto-learn's synthesis for accuracy
- thinking-dimensions: Auto-learn findings → D1 (perspective matrix) enriched with latest data
Anti-Patterns
- Learning for learning's sake (must serve the current task)
- Trusting the first search result (cross-verify required)
- Storing without verifying (false knowledge is worse than no knowledge)
- Re-learning what memory-lora already has at score ≥0.8
Success Metric
Number of times a stored learning is RECALLED and APPLIED in a later task. A learning that's stored but never used = wasted token budget. A learning that prevents one wrong answer = 10x ROI.