Instinct System

1. OBSERVE
   During normal work, notice something that might be a pattern.
   "This handler is internal sealed class. Is that the convention?"

2. HYPOTHESIZE
   Create an instinct with initial confidence 0.3.
   Write to .claude/instincts.md:
   - Use `sealed` on all handler classes | confidence: 0.3 | seen: 1 | last: 2025-07-15

3. SEEK CONFIRMATION
   When working in related areas, actively check for the pattern.
   Open 2-3 other handlers. Do they all use `internal sealed class`?

4. CONFIRM or DISCONFIRM
   - If confirmed: increment `seen`, raise confidence per the adjustment rules
   - If contradicted: halve confidence, note the exception
   - If mixed: hold confidence steady, note the split

5. PROMOTE
   At 0.9, present to the user for promotion to MEMORY.md.
   "I've observed [pattern] across [N] instances. Should I add this
   as a permanent rule to MEMORY.md?"

# Project Instincts

## Code Style [0.7]
- Use `sealed` on all handler classes | confidence: 0.8 | seen: 5 | last: 2025-07-15
- Prefix private fields with underscore | confidence: 0.5 | seen: 2 | last: 2025-07-14

## Architecture [0.6]
- Feature folders use singular names (Order, not Orders) | confidence: 0.7 | seen: 4 | last: 2025-07-15

## Naming [0.5]
- Command classes end in Command, not Request | confidence: 0.5 | seen: 3 | last: 2025-07-15

CONFIRMATION TRACK:
  First observation         → 0.3   (hypothesis formed)
  Second confirmation       → 0.5   (pattern emerging)
  Third confirmation        → 0.7   (likely convention)
  Fourth+ confirmation      → 0.8   (strong convention)
  Fifth+ with zero contradictions → 0.9 (promotion candidate)

CONTRADICTION HANDLING:
  Any contradiction          → halve current confidence
  Example: instinct at 0.7, contradicted → drops to 0.35
  Two contradictions in a row → drop to 0.1 (effectively dead)

USER OVERRIDE:
  User explicitly confirms   → jump to 0.8
  User explicitly corrects   → drop to 0.0 (remove the instinct)
  User says "sometimes"      → cap at 0.5 (conditional pattern)

STALENESS:
  No new observations for 10+ sessions → flag for review
  Contradicted and not re-confirmed for 5 sessions → remove

0.0 - 0.2  → IGNORE — insufficient evidence, do not mention
0.3 - 0.4  → NOTE — record internally, do not apply
0.5 - 0.6  → MENTION — "I notice this project may use [pattern]. Should I follow it?"
0.7 - 0.8  → FOLLOW — apply by default, mention when first applied
0.9        → PROMOTE — offer to add to MEMORY.md as a permanent rule

ACTIVE SEEKING PROTOCOL:
1. Form instinct: "This project uses Result<T> for handler returns"
2. Before next code generation, check 2-3 existing handlers:
   → Use find_symbol to locate other handlers
   → Use get_public_api to check their return types
3. Count matches and mismatches
4. Adjust confidence based on findings
5. Update .claude/instincts.md with new count and confidence

EXAMPLE:
  New instinct: "Handlers return Result<T>" at 0.3
  → Check CreateOrderHandler: returns Result<OrderResponse> ✓
  → Check GetProductHandler: returns Result<ProductResponse> ✓
  → Check DeleteUserHandler: returns Task<IResult> ✗
  Result: 2/3 match → raise to 0.5, note the exception
  Updated: Handlers return Result<T> | confidence: 0.5 | seen: 3 | last: 2025-07-15
           Exception: DeleteUserHandler uses Task<IResult>

PROMOTION STEPS:
1. Present the evidence to the user:
   "I've observed [pattern] across [N] instances with zero contradictions:
    - CreateOrderHandler: ✓
    - UpdateOrderHandler: ✓
    - DeleteOrderHandler: ✓
    - GetProductHandler: ✓
    - CreateProductHandler: ✓
    Should I add this as a permanent rule to MEMORY.md?"

2. If user approves:
   - Add to MEMORY.md under the appropriate category
   - Remove from .claude/instincts.md
   - Format as a clear, generalized rule (use self-correction-loop generalization)

3. If user declines:
   - Cap confidence at 0.8
   - Note: "User reviewed, chose not to promote"
   - Keep in instincts for continued reference

4. If user provides context:
   - "That's only for command handlers, not query handlers"
   - Adjust the instinct to be more specific
   - Reset confidence to 0.5 (narrowed scope needs re-confirmation)