Claude Improve

/improve

A self-improving retrospective skill for Claude Code. One command makes your AI agents get better after every conversation.

`/improve` reviews your conversations, detects patterns in your feedback, cross-references against your config files, and suggests targeted improvements — presented as a ranked findings list, applied only with your approval.

The model stays the same. The instructions get smarter.

**v4.0.0** — Two-level learnings (global preferences + per-project state), compact findings-list presentation by default, simplified scope question, and a settled-pattern promotion rule so learnings graduate into your config. See [Releases](https://github.com/TerenceBristol/claude-improve/releases) for the full changelog.

---

Quick Start

**1. Clone this repo**

git clone https://github.com/TerenceBristol/claude-improve.git

**2. Copy the skill to your Claude Code commands**

cp claude-improve/improve.md ~/.claude/commands/improve.md

**3. Run it in any Claude Code project**

/improve

That's it. The skill works on any project with configuration files (CLAUDE.md, skills, agents, memory, frameworks). No dependencies, no setup, no API keys.

---

Updating

Pull the latest and re-copy:

cd ~/path/to/claude-improve && git pull
cp improve.md ~/.claude/commands/improve.md

---

Files Created

`/improve` maintains two learnings files automatically:

• ~/.claude/improve-learnings.md (global) — Cross-project patterns about how you like the skill and rule-writing to work: presentation preferences, scope habits, modify signals. Read at the start of every run in every project.
• ~/.claude/projects//improve-learnings.md (per project) — That project's distilled patterns, deferral counters, and a date-keyed run log. Lives in Claude Code's own per-project data directory under your home folder, so it never touches your repo. Note: is the project's absolute path with every