# Capture corrections and update your CLAUDE.md memory

Claude Code forgets your corrections between sessions, so you repeat the same guidance again and again. claude-reflect captures those corrections automatically and lets you review and apply them to your CLAUDE.md memory files with one command.

## Install

```bash
npx skillstore add bayramannakov/claude-reflect
```

## Metadata

- Status: approved
- Slug: bayramannakov-claude-reflect
- Skillstore revision: r2
- Version status: missing
- Tree hash: 86e6fac3dec9f16db3e5a6537e605c095a28f6203160124c5cd4fd0ba32221b6
- Author: BayramAnnakov
- GitHub username: BayramAnnakov
- License: MIT
- Repository: https://github.com/BayramAnnakov/claude-reflect/tree/main/
- Ref: 26ee580b1dc4b5aed6b30c1dd62b327556bac8d8
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: filesystem, external\_commands, network
- Quality score: 79
- Quality tier: bronze
- Public page: https://skillstore.pages.dev/skills/bayramannakov-claude-reflect
- Manifest: https://skillstore.pages.dev/api/skills/bayramannakov-claude-reflect/manifest

## Capabilities

- Detects correction patterns in your prompts, such as 'no, use X' or 'actually', using a hooks-based capture stage.
- Queues detected learnings to a local JSON file for later review.
- Provides a /reflect command that reviews queued learnings with human approval before writing.
- Routes approved learnings to the correct memory target, including global, project, and rule files.
- Scans past session files for missed corrections with /reflect --scan-history.
- Suggests reusable skill candidates from repeating patterns with /reflect-skills.

## Use Cases

- Stop repeating the same corrections: A developer who keeps telling Claude Code the same project conventions can capture those corrections once and apply them to the project CLAUDE.md file.
- Keep team conventions in sync: A team lead can review queued learnings before they are written, so shared CLAUDE.md files stay accurate and reviewed.
- Recover missed learnings from past work: A user returning to a project can scan past sessions to find corrections that were never saved and add them to memory.

## Prompt Templates

### Review queued learnings

```
Run /reflect and show me the learnings that are queued before applying any of them.
```

### Preview changes safely

```
Run /reflect --dry-run so I can see what would change in my CLAUDE.md files without writing anything.
```

### Scan history for missed learnings

```
Run /reflect --scan-history and list corrections from past sessions that were never saved.
```

### Discover reusable skills

```
Run /reflect-skills over the last 30 days and propose skill candidates from repeating patterns, but do not create any files yet.
```

## Limitations

- It writes only to Markdown memory files; it does not change your code or run your tools.
- Correction detection depends on recognizable phrasing and may miss subtle guidance.
- The semantic analysis step requires the Claude CLI to be available.
- It is designed for Claude Code and its CLAUDE.md hierarchy, not general note-taking.

## Best Practices

- Review learnings with /reflect --dry-run before applying them so you can see the exact changes.
- Run /reflect at natural stopping points, such as after finishing a feature or before context compaction.
- Choose the narrowest memory target that fits, using project files for project rules and global files only for broad patterns.

## Anti Patterns

- Do not apply every queued learning without reading it, because low-value corrections can clutter your memory files.
- Do not store secrets, tokens, or personal data in captured learnings, since memory files persist across sessions.
- Do not rely on capture alone to remember important rules; run /reflect so learnings are actually written.

## Security Audit

- Audited at: 2026-07-21T10:00:40.065\+00:00
- Summary: claude-reflect is a documentation-only Claude Code plugin corpus \(8 Markdown/text files, no executable scripts in scope\). All 91 static findings were adjudicated as false positives: the \`~/.claude/\*\` references are the plugin's own documented memory-hierarchy and queue storage, the two CRITICAL \`rm -rf\` hits are user-facing cache-clearing instructions scoped to a specific named plugin-cache subdirectory in the README troubleshooting section, the 'backtick execution' hits are Markdown inline-code/code-fence syntax, and the hardcoded URLs point to the project's own public repository. No prompt-injection, data-exfiltration, or business-logic-abuse intent was found.

## Stats

- Views: 231
- Downloads: 31
- Favorites: 2
- Popularity score: 0