continuous-learning-v2

Continuous Learning v2 (instinct-based)

A finer-grained alternative to continuous-learning: instead of evaluating a whole session once at the end and writing out full skill files, this approach observes every tool call via hooks and accumulates small, atomic, confidence-scored "instincts" that later get clustered into skills, commands, or agents. See continuous-learning for when the simpler whole-session approach is enough — don't reach for this added complexity unless you actually need per-behavior confidence tracking or project/global separation.

Why hooks instead of a Stop hook

A Stop hook only fires once, on sessions that end normally, and only gets one pass over the whole transcript. PreToolUse/PostToolUse hooks fire on every single tool call, deterministically — so no candidate pattern is missed because the session ended unusually, and observation happens incrementally rather than in one large end-of-session pass.

The instinct model

An instinct is one atomic, learned behavior — one trigger, one action — not a full skill:

---
id: prefer-explicit-error-returns
trigger: "when writing a function that can fail"
confidence: 0.7
domain: "code-style"
source: "session-observation"
scope: project
project_id: "a1b2c3d4e5f6"
---

# Prefer Explicit Error Returns

## Action
Return an explicit error value rather than panicking, for any function
whose failure is an expected, recoverable condition.

## Evidence
- User corrected a panic-based approach to an error-return approach twice
- Pattern held across three subsequent sessions without correction

Properties that make this different from a whole extracted skill:

• Atomic — one trigger, one action, easy to evaluate independently
• Confidence-weighted — a number, not a boolean, that can move up or down as evidence accumulates
• Domain-tagged — code-style, testing, git, debugging, etc.
• Evidence-backed — records what observations produced it, so a human reviewing it later can see why it exists
• Scope-aware — project by default, promotable to global