Making Claude Code remember it was wrong

What it does\n\nclaude-reflect is a Claude Code plugin that listens for corrections ("no, use X not Y") and positive feedback during sessions, queues them, and—after human review—persists them into CLAUDE.md or AGENTS.md. It also mines your session history for repeating tasks and suggests turning them into reusable slash-commands under .claude/commands/.\n\nThe interesting bit\n\nIt runs a two-stage pipeline: lightweight regex hooks catch potential learnings in real time, but an AI semantic filter validates them later during /reflect, which means it understands corrections in any language, not just English regex matches. Corrections made while using a specific skill can be routed back into that skill’s own markdown file, so /deploy learns from its mistakes.\n\nKey highlights\n- Captures corrections, praise, and explicit "remember:" markers automatically via session hooks, but nothing gets written until you approve it.\n- Semantic pattern detection groups similar requests across different phrasings to suggest skill candidates.\n- Generates draft skill files with guardrails, context blocks, and step-by-step workflows.\n- Syncs learnings globally, per-project, or per-subdirectory, and writes to AGENTS.md for compatibility with Codex, Cursor, Aider, and others.\n- Ships with 160 passing tests and supports macOS, Linux, and Windows natively.\n\nCaveats\n- Upgrading from v2.0.x or earlier can hit Claude Code caching bugs that require manually clearing plugin directories and restarting; the README links to two open Anthropic issues about cache invalidation.\n\nVerdict\nIf you spend enough time in Claude Code to accumulate a personal style guide of corrections, this turns those moments into persistent instructions. If you don’t use Claude Code or rarely correct it, this is just a plugin watching an empty room.