What problem does it solve? Corrections, workarounds, and preferences discovered during Claude Code sessions are easily lost, so the same mistakes repeat. This Skill scans recent memory files, git history, and session context, then turns those observations into concrete, evidence-backed improvements to CLAUDE.md, skills, hooks, and memory files via a reviewable pull request. ## Core Features & Use Cases - Multi-source observation gathering: Reads auto memory files under ~/.claude/projects/*/memory/, recent git commits, and current session friction within a configurable lookback window (default 8 hours). - Categorized improvement proposals: Groups findings into coding conventions, workflow preferences, tool usage, missing skills, hook improvements, and memory gaps, each mapped to a target file. - Approval-gated changes: Presents every proposed change with cited evidence and requires explicit user approval before editing any file. - PR-based delivery: Applies approved changes on a dated branch and opens a PR summarizing each change with its motivating observation. - Use Case: After a week of sessions where you repeatedly corrected commit message style and TypeScript typing, run the Skill to get a PR that codifies those rules into CLAUDE.md. ## Quick Start Ask Claude to run the claude-doctor skill to review the last 8 hours of learnings and propose configuration improvements.
