# Learning-Loop Skill for Claude Code

A two-mode Claude Code skill that catches what your sessions teach you — failed attempts, user corrections, recurring failure modes, judgment shifts — and writes them to the right files before `/clear` or context compaction destroys the details.

> **Heads up: this is a personal skill, published in case it's useful.** It's shaped by one workflow — a root `~/.claude/CLAUDE.md`, per-project `CLAUDE.md` files, a `MEMORY.md`, a personal Judgment Ledger, and Every's [`/ce-compound`](#what-is-ce-compound) for code-level capture. If your setup looks roughly like that, the routing will land where you'd expect. If it doesn't, you'll want to read [SKILL.md](SKILL.md) and adapt the destinations.

## The Problem

Claude Code sessions accumulate valuable signal — a hypothesis that turned out to be right, a fix you confirmed, a workflow rule you broke and want to encode, a corrected assumption, a recurring mistake you keep making across projects. Most of that gets lost the moment context compacts or you hit `/clear`.

Claude Code's built-in auto-memory captures quick facts, but it doesn't: - Distinguish a one-off observation from a pattern you've now hit five times - Apply quality gates (would this help next time? did you actually verify the fix?) - Route a process-level rule into `CLAUDE.md` vs. a factual recall into `MEMORY.md` vs. a judgment shift into a content ledger - Surface recurring failure modes that need a structural fix rather than another note

Learning-loop is the structured pass on top of that — invoked explicitly, run by sub-agents so it doesn't eat your main context, and gated so it doesn't pollute your docs with noise.

## What It Does

| Mode | When you run it | What it does | |---|---|---| | **`/learning-loop scan`** | Mid-session, before compaction or `/clear` | Spawns a sub-agent that reads your conversation, extracts raw signals (failed attempts, hypotheses, user corrections, process observations), and writes them to `