LearnWatch

The problem it solves

Claude Code fixes the same classes of bugs repeatedly across sessions because it has no persistent memory of what went wrong before. LearnWatch breaks this cycle: every fix you confirm gets evaluated, and the lessons that survive review are written into your permanent coding standards — so Claude Code learns from your past mistakes.

How graduation works

1. Enqueue — When you fix a bug or notice a pattern, add a hypothesis to LearnWatch via the Learning tab or the Claude Code hook that fires automatically on session end.
2. Judge — LearnWatch uses your Anthropic key to run a lightweight Haiku evaluation: is this lesson novel, specific, and actionable? Vague or duplicate hypotheses are rejected.
3. Graduate — Approved lessons are written to one of four artifact types: lesson (LESSONS.md), anti-pattern (anti-patterns.md), design pattern (design-patterns/), or memory (MEMORY.md). These files are read by Claude Code at the start of each session.

Artifact types

• Lesson — A specific coding rule derived from a real bug. Example: "Never use bare except: — catch specific exception types."
• Anti-pattern — A recurring mistake with a name, symptom, and fix. Appears in Claude Code's anti-patterns registry.
• Design pattern — A positive pattern worth repeating. Counterpart to anti-patterns.
• Memory — Project-specific or user-specific context that should persist across sessions.

Costs

The judge step uses your Anthropic API key and calls Claude Haiku — typically under $0.001 per hypothesis. Processing a backlog of 20 hypotheses costs roughly $0.02. LearnWatch never runs evaluation automatically in the background; it only runs when you trigger it or when your Claude Code session ends.

Using the Learning tab

1. Open Learning in the Sentinel sidebar.
2. The stats panel shows pending / graduated / rejected counts.
3. Click + Enqueue hypothesis to add a new lesson candidate manually.
4. Recently graduated lessons appear in the table with their artifact type and graduation date.