bayramannakov/claude-reflect

Overview

This skill captures user corrections, preferences, and positive feedback during chat sessions and helps you persist them into CLAUDE.md and related learning files. It runs automatic hooks to queue likely learnings and provides a manual /reflect workflow so you review and apply changes with human judgment. The system reduces repeated corrections and preserves project- or global-level guidance for future sessions.

How this skill works

Hooks monitor conversation patterns and high-confidence correction phrases, queueing candidate learnings to a local queue file. When you run /reflect, the skill presents queued items, suggests scopes (global, project, or command-level), and allows you to accept, edit, or discard each change. Additional commands let you scan session history, preview edits with --dry-run, discover recurring patterns as skill candidates, or clear the queue.

When to use it

• After you make a correction you expect to matter in future sessions ("no, use X")
• When you explicitly ask the assistant to remember something for later
• After finishing a feature or meaningful work unit to capture conventions
• When context is about to be compacted and the queue contains items
• Periodically, to review accumulated learnings and keep CLAUDE.md current

Best practices

• Run /reflect regularly rather than letting many learnings accumulate
• Prefer reviewing and editing suggested entries before applying them
• Use --dry-run to preview changes before committing to documentation
• Choose the correct scope (global vs project vs skill) to avoid noise
• Use /reflect-skills to surface repeated corrections that deserve a formal skill

Example use cases

• You tell the assistant to prefer a different model for reasoning; capture and persist that preference globally.
• A project has a naming convention you corrected; add it to the project CLAUDE.md so future sessions follow it.
• During a skill execution you correct output behavior; convert repeated corrections into a commands/*.md improvement.
• Use --scan-history to find missed corrections from previous sessions and bring them into the queue.
• Preview all queued learnings with /view-queue and then apply only the ones that matter using /reflect --dry-run.