Learn

Learn Skill - Continuous Learning System

Purpose

Capture reusable patterns discovered during sessions as "instincts" with confidence scoring. Instincts start weak (0.3) and grow stronger with repeated confirmation (max 0.9).

Trigger

• User types /learn followed by a pattern description
• User types /learn with no args (auto-detect from current session context)

Algorithm

1. Parse input: Extract the pattern description and optional tags from user input
2. Read instincts file: Load ~/.claude/learnings/instincts.json
   • If file is empty or missing, initialize with default structure
3. Check for duplicates: Compare new pattern against existing instincts
   • Fuzzy match: if 60%+ of significant words (4+ chars) overlap, treat as existing
   • If match found: increment confidence by 0.2 (cap at 0.9), add new evidence entry, update last_confirmed
   • If no match: create new instinct with confidence 0.3
4. Generate instinct object (for new patterns):

   {
     "id": "inst_NNN",
     "pattern": "<pattern description>",
     "confidence": 0.3,
     "evidence": ["session_YYYY-MM-DD: <context>"],
     "created": "YYYY-MM-DD",
     "last_confirmed": "YYYY-MM-DD",
     "tags": ["<auto-detected-or-user-provided>"],
     "source": "user|auto",
     "evolved_to": null
   }
   

5. Save: Write updated instincts.json, increment metadata.total_learned
6. Display: Show the saved/updated instinct to the user in a clear format

ID Generation

• Format: inst_NNN where NNN is zero-padded 3-digit sequential number
• Read existing instincts, find max ID number, increment by 1
• First instinct: inst_001

Tag Auto-Detection

If user does not provide tags, infer from pattern text using these keyword groups:

Assign up to 3 tags. If no keywords match primary categories, try secondary:

If still no match, use ["general"] (last resort).

Fuzzy Match Logic

significant_words(text) = [word.lower() for word in text.split() if len(word) >= 4]
# Jaccard similarity (intersection/union) is symmetric and handles
# different-length patterns better than precision-style overlap
jaccard(a, b) = len(set(significant_words(a)) & set(significant_words(b))) / len(set(significant_words(a)) | set(significant_words(b)))
is_duplicate = jaccard >= 0.6

Confidence Scale

Formula: new_confidence = min(current_confidence + 0.2, 0.9) Contradiction: new_confidence = max(current_confidence - 0.2, 0.0) -> archive at 0.0

Output Format

INSTINCT CAPTURED

ID:         inst_NNN
Pattern:    <description>
Confidence: 0.X [NEW | +0.2 CONFIRMED]
Tags:       tag1, tag2
Evidence:   N entries

Tip: Use /evolve when you have 3+ instincts at 0.7+ to create reusable skills.

Corruption Detection and Recovery

Detection criteria: instincts.json is corrupted if:

• Invalid JSON syntax
• Missing "version" field
• Missing "instincts" array
• File size > 10MB

Recovery:

1. Back up as instincts.json.YYYYMMDD_HHMMSS.bak (timestamped, not overwrite)
2. Keep max 3 backups (delete oldest)
3. Reinitialize with fresh schema
4. Log: "Recovered from corrupted instincts.json"

Error Handling

• If no pattern text provided and no session context available, ask user to describe the pattern
• Maximum 500 instincts in file; if exceeded, archive instincts with confidence < 0.3 that are older than 30 days