Instinct System

Instinct System

One learning system, three tiers. Every signal Claude receives during work — an observed pattern, a user correction, a surprising discovery — routes to exactly one store.

Core Principles

1. Three tiers, one routing decision — Instincts are unconfirmed hypotheses (.claude/instincts.md). Corrections are user-confirmed rules (MEMORY.md). Discoveries are insights that explain the world (.claude/learning-log.md). Rules prescribe behavior; insights describe it; instincts are rules-in-waiting. Never mix the tiers — a hypothesis in MEMORY.md pollutes permanent knowledge, and a confirmed rule left as an instinct gets forgotten.

1. Instincts are hypotheses, not rules — An instinct starts as a guess from a single observation and has no authority until confirmed. "One handler uses sealed" is an instinct at 0.3; "all 12 handlers use sealed" is a rule at 0.9. Confidence (0.3–0.9) drives behavior: note at 0.3, mention at 0.5, follow at 0.7, promote at 0.9.

1. A user correction is a confirmed instinct at full confidence — When the user corrects you, skip the confirmation cycle entirely. Generalize the lesson, capture it immediately in MEMORY.md with the "why", and confirm what was captured. A correction costs the user 30 seconds today and saves hours across all future sessions — losing one is the most expensive mistake this system can make.

1. Project-scoped, never global — What holds in one codebase may be wrong in another. Instincts live per-project; transfers between projects go through export/import with confidence decay, never at full confidence.

1. Review at session start, prune periodically — Read MEMORY.md and load instincts at 0.7+ before writing any code; scan recent log entries for the working area. Knowledge captured but never reviewed is wasted effort. Audit all three stores when they bloat — stale instincts, duplicate rules, and unreviewed logs defeat the purpose.

Patterns

Tier Routing

Instinct Lifecycle (Observe-Hypothesize-Confirm)

1. OBSERVE     — "This handler is internal sealed class. Convention?"
2. HYPOTHESIZE — Write to .claude/instincts.md at confidence 0.3:
                 - Use `sealed` on handler classes | confidence: 0.3 | seen: 1 | last: 2026-06-12
3. SEEK        — Actively check 2-3 related files (find_symbol, get_public_api).
                 Passive observation is not enough.
4. ADJUST      — Confirmed: raise per the adjustment track.
                 Contradicted: halve confidence, note the exception.
                 Mixed: hold steady, note the split.
5. PROMOTE     — At 0.9, present evidence and offer promotion to MEMORY.md.

Instinct Storage Format

.claude/instincts.md, grouped by category with structured metadata:

# Project Instincts

## Code Style [0.7]
- Use `sealed` on all handler classes | confidence: 0.8 | seen: 5 | last: 2026-06-12
- Prefix private fields with underscore | confidence: 0.5 | seen: 2 | last: 2026-06-10

## Architecture [0.7]
- Feature folders use singular names | confidence: 0.7 | seen: 4 | last: 2026-06-12

Category header [0.7] = average confidence. Standard categories: Code Style, Architecture, Naming, Testing, Data Access, API Design, Configuration, Performance, Tooling.

Confidence Adjustment Rules

No ad-hoc scoring. Follow this track precisely:

CONFIRMATION:  1st observation → 0.3 | 2nd → 0.5 | 3rd → 0.7 | 4th → 0.8
               5th+ with zero contradictions → 0.9 (promotion candidate)

CONTRADICTION: Any contradiction → halve current confidence (0.7 → 0.35)
               Two in a row → 0.1 (effectively dead)

USER OVERRIDE: Explicit confirm → 0.8 | Explicit correct → 0.0 (remove,
               capture in MEMORY.md instead) | "Sometimes" → cap at 0.5

STALENESS:     No observations for 10+ sessions → flag for review
               Contradicted, unreconfirmed for 5 sessions → remove

Acting on Instincts by Confidence

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 — "This project may use [pattern]. Follow it?"
0.7–0.8 → FOLLOW  — apply by default, mention on first use
0.9     → PROMOTE — offer to add to MEMORY.md as a permanent rule

Never silently apply an instinct below 0.7.

Correction Capture Flow (Tier 2)

When the user corrects your output ("no, use X instead", "we don't do it that way", "always/never do X", "remember this"):

1. DETECT      — Recognize the correction signal.
2. ACKNOWLEDGE — "Got it — HybridCache instead of IMemoryCache."
3. GENERALIZE  — Specific: "Don't use IMemoryCache in the Orders endpoint"
                 General:  "Always use HybridCache over IMemoryCache —
                            stampede protection + L1/L2 out of the box."
                 Ask: is this specific to this file? This layer? What's the
                 underlying principle? Store the broadest correct statement.
4. CHECK       — Scan MEMORY.md for overlap. Update an existing rule rather
                 than adding a near-duplicate.
5. STORE       — Write under the right category, one line, rationale after
                 the dash. If a matching instinct exists, remove it — it
                 just graduated.
6. CONFIRM     — "Added to Memory > Data Access: HybridCache over IMemoryCache."

MEMORY.md format — same categories as instincts, one actionable rule per line:

## Data Access
- Always use HybridCache over IMemoryCache — stampede protection + L1/L2
- Never use repository pattern over EF Core — use DbContext directly

Discovery Logging (Tier 3)

Log organic findings to .claude/learning-log.md the moment they occur — a 2-line entry written immediately beats a paragraph reconstructed later. Entry format:

## 2026-06-12 | Gotcha | MassTransit Consumer Registration Order Matters
Multiple consumers for one message type run in registration order; if the
first throws, the rest are skipped. Caused missed audit events.
**Files:** `src/Shared/Extensions/MassTransitConfig.cs:15-30`
**Resolution:** Independent consumer endpoints or the retry filter.

Log when (and only when) you hit one of these triggers:

Bug Root Cause        — the cause was NOT where the error appeared
Architecture Decision — discovered WHY something is built a certain way
Gotcha                — framework/library behaved surprisingly
Performance Discovery — unexpected perf behavior, with the measurement
Pattern Found         — reusable codebase pattern worth remembering
External Service      — third-party API behaves differently than documented

Those six trigger names are also the category vocabulary — use them verbatim in entries. Routine changes with nothing surprising do not get logged.

Session-Start Loading

1. Read MEMORY.md — apply rules proactively, don't wait to be reminded
2. Read .claude/instincts.md — load 0.7+ as defaults, note 0.5–0.6
3. Scan recent learning-log entries for the area you're working in
4. Flag stale instincts (10+ sessions without observation) for review

Modes

Three operations on the instinct store, invoked by name or trigger phrase.