insight-collector

Insight Collector

Hybrid learning system: manual insights + automatic pattern learning.

System Overview

Manual Insights (★ blocks) → .caw/insights/
Automatic Instincts (patterns) → .caw/instincts/
Evolution (≥0.6 confidence) → .caw/evolved/

Part 1: Manual Insight Capture

Protocol

1. Generate: Display ★ Insight block with 2-3 key points
2. Save: Immediately write to .caw/insights/{YYYYMMDD}-{slug}.md
3. Confirm: 💡 Insight saved: [title]

When to Generate

• Implementation discovery (useful patterns)
• Problem solution (lessons learned)
• Best practice (project-specific)
• Gotcha/Pitfall (traps to avoid)
• Architecture decision (design rationale)

Part 2: Automatic Pattern Learning

How It Works

Tool Usage → Observation (hooks/observe.py) → Pattern Detection → Instinct

Detected Patterns

Pattern	Example	Instinct
Tool Sequence	Grep → Edit → Grep	"Verify with Grep after Edit"
Error Recovery	Edit fails → retry	"Adjust parameters on retry"
Tool Preference	80% Grep over search	"Prefer Grep for code search"
Workflow	Same 3-tool sequence	"Standard modification workflow"

Confidence Scoring

Evidence	Confidence
1-2 obs	0.3
3-5 obs	0.5
6-10 obs	0.7
11+ obs	0.9 max

Changes: Confirming +0.05, Contradicting -0.10, Weekly decay -0.02

CLI Commands

python3 scripts/instinct-cli.py analyze [--incremental|--full]
python3 scripts/instinct-cli.py list|show|promote|demote|decay|stats
python3 scripts/instinct-cli.py export|import
python3 scripts/instinct-cli.py dashboard

Part 3: Evolution

Condition	Evolution
Confidence ≥0.6	Eligible
User-triggered (3+ steps)	→ Command
Auto-applicable	→ Skill
Complex reasoning	→ Agent

Use /crew:manage evolve to preview and generate.

Directory Structure

.caw/
├── insights/        # Manual (Part 1)
├── instincts/       # Automatic (Part 2)
│   ├── index.json
│   └── personal/
├── observations/    # Raw data
└── evolved/         # Part 3: commands/skills/agents

Tag Generation

Content Pattern	Tag
auth, jwt, login	#authentication
security, xss	#security
performance, cache	#performance
test, mock	#testing
api, endpoint	#api

Boundaries

Will: Save insights immediately, observe patterns, generate instincts, auto-tag Won't: Modify content arbitrarily, overwrite without confirmation, auto-evolve low-confidence