TL;DR
Claude Reflect is a Claude Code plugin that captures user corrections and positive feedback during sessions, queues them for review, and can sync accepted items into CLAUDE.md and AGENTS.md. It uses hooks to auto-detect patterns, assigns confidence scores, and provides a manual /reflect workflow for review, deduplication and historical scans.
What happened
A new open-source plugin called claude-reflect adds a lightweight, two-stage learning loop to Claude Code. Capture hooks run automatically while you interact with Claude, spotting correction phrases (for example "use X not Y" or explicit "remember:" markers) and some praise patterns, then queue those findings with a confidence score. When you run the /reflect command you get a summary table of queued learnings and can apply selected items to global (~/.claude/CLAUDE.md) or project-local (./CLAUDE.md) configuration, and optionally AGENTS.md if present. The tool includes commands to scan past sessions, preview changes, review queue details, deduplicate semantically similar entries, and skip queued items. Installation is via the Claude plugin marketplace and requires a restart of Claude Code; prerequisites include the Claude Code CLI, jq and Python 3. The repository is published under an MIT license and accepts pull requests.
Why it matters
- Preserves developer corrections and preferences so Claude can recall them across sessions instead of repeating the same adjustments.
- Provides a human-reviewed bridge from ephemeral chat corrections into persistent project or global configuration files.
- Reduces repetitive manual updates to CLAUDE.md and AGENTS.md through automated detection and deduplication tools.
- Supports auditing: each captured learning carries a confidence score and requires human review before being applied.
Key facts
- Plugin name: claude-reflect; repository: BayramAnnakov/claude-reflect on GitHub.
- Install via Claude plugin marketplace: `claude plugin marketplace add bayramannakov/claude-reflect` and `claude plugin install claude-reflect@claude-reflect-marketplace`.
- Restart Claude Code after installation so hooks auto-configure and commands become available.
- Main command: /reflect — processes queued learnings with human review; flags include –scan-history, –dry-run, –targets, –review, and –dedupe.
- Automatic capture uses hook scripts (capture-learning.sh, check-learnings.sh, post-commit-reminder.sh) to queue detected corrections and positive feedback.
- Each captured learning is scored with a confidence value (0.60–0.95) based on pattern strength; explicit markers like "remember:" carry highest confidence.
- Supports syncing accepted learnings to global CLAUDE.md (~/.claude/CLAUDE.md), project CLAUDE.md (./CLAUDE.md), and AGENTS.md when present.
- Provides tools for semantic deduplication, merging, replacing or skipping similar existing entries before updating files.
- Repository lists MIT license; codebase includes commands, hooks, scripts and a SKILL.md for context awareness.
- GitHub repo metadata shown in source: 6 stars and 0 forks at time of capture.
What to watch next
- Adoption and community feedback about how reliably the pattern-detection works in varied workflows (not confirmed in the source).
- How the plugin handles privacy or sensitive data captured from sessions, and any guidance on filtering such content (not confirmed in the source).
- Updates to compatibility with future Claude Code versions or additional agent integrations (not confirmed in the source).
Quick glossary
- Claude Code: A developer-focused interface or client for interacting with Anthropic's Claude models (term used generically in the source).
- CLAUDE.md: A configuration or preference file where persistent instructions and conventions for Claude are stored at global or project scope.
- Hook: A scripted trigger that runs automatically in response to events (for example, each prompt or after git commits) to capture data or perform checks.
- Agent (AGENTS.md): An auxiliary tool or integration (Codex, Cursor, Aider, etc.) documented in AGENTS.md; used here as a target for syncing applicable learnings.
Reader FAQ
How do I install claude-reflect?
Add the plugin marketplace entry and install via the Claude CLI: `claude plugin marketplace add bayramannakov/claude-reflect` then `claude plugin install claude-reflect@claude-reflect-marketplace`, and restart Claude Code.
What does the /reflect command do?
/reflect shows queued learnings with confidence scores, lets you review and select items, and applies accepted entries to CLAUDE.md or AGENTS.md; it also supports scanning history and deduplication flags.
Does claude-reflect automatically edit my files without review?
No — capture is automatic, but applying learnings to files is manual and done via /reflect, which offers review and selection options.
Can it scan past sessions and clean duplicates?
Yes. There are commands to scan history (`/reflect –scan-history`) and to find and consolidate semantically similar entries (`/reflect –dedupe`).
claude-reflect A self-learning system for Claude Code that captures corrections, positive feedback, and preferences — then syncs them to CLAUDE.md and AGENTS.md. What it does When you correct Claude Code…
Sources
- Show HN: Claude Reflect – Auto-turn Claude corrections into project config
- Claude Code: Best practices for agentic coding
- Claude Code settings, commands and agents for vibe coding
- Claude Code settings – Claude Code Docs
Related posts
- Neural Networks: Zero to Hero — Andrej Karpathy’s from-scratch course on LLMs
- Grok’s ‘apologies’ are unreliable in non-consensual image controversy
- SQLNet: A Twitter-like social network where you use SQL to interact