self-improve
Captures user corrections and improvements during interactions, logs them, finds recurring patterns, and prescribes preventive fixes like CLAUDE.md rules, skills, scripts, or MCP tools to avoid repeating mistakes.
README
self-improve
한국어: README.ko.md
A Claude Code plugin that turns your corrections into lasting improvements.
The moment you fix or redo Claude's output is the most valuable data about the gap between what you wanted and what Claude produced. Today that data evaporates when the conversation ends. self-improve captures those moments, finds recurring patterns, and prescribes recurrence-prevention fixes — CLAUDE.md rules, skills, scripts, or MCP tools.
Install
/plugin marketplace add Feynman520/d06-p01-self-improve
/plugin install self-improve@self-improve
Requires Node.js >= 18. No npm install needed — the plugin has zero runtime dependencies.
How it works
Hybrid logging — Claude detects and logs automatically (always-on instructions shipped with the MCP server), and you can log manually with /self-improve-log.
Claude watches for 4 moments:
- You correct or redo a result ("no, not that", "do it again")
- You declare a future rule ("from now on, always ...")
- A tool/command fails and Claude works around it
- The same manual work repeats yet again
Noise filter: brand-new requests and one-off taste tweaks are not logged. The bar is the one-sentence test — if no reusable lesson can be stated in one sentence, nothing is recorded. After each log, Claude prints one line (📝 개선점 기록: "..." (category)); tell it to undo and the record is deleted.
Categories (fixed)
| Category | Meaning |
|---|---|
| 오해 | Misread your request |
| 지식 부족 | Missing user/environment-specific fact |
| 도구 실패 | Tool or command failure |
| 환경 문제 | Config / path / permission issue |
| 반복 비효율 | Repeated manual work (automation candidate) |
MCP tools
| Tool | Purpose |
|---|---|
log |
Save a record; auto-links similar past records (related) — repetition is evidence of a pattern |
search |
Filter records by keyword / category / project / status |
stats |
Aggregate by category / project / month |
resolve |
Mark records fixed, with the applied prescription (bulk ids) |
Skills
/self-improve-log— log an improvement point from the current conversation right now./self-improve-report— periodic report: groups open records into patterns, then climbs the prescription ladder (CLAUDE.md rule → skill → script/library → MCP; always the lightest fix that works), sorted by impact-per-cost. Warns at the top when a resolved problem re-occurred (⚠️ prescription failed). Saved toreports/YYYY-MM-DD-보고.md.
Recommended cadence: run /self-improve-report after ~10 records or every 2–4 weeks.
Storage
One record = one markdown file (frontmatter + 4 sections), fully local.
- Default:
.self-improve/in the current project (records/,reports/) - Override: set the
SELF_IMPROVE_DIRenvironment variable to a central folder
---
id: 2026-07-07-001
date: 2026-07-07
project: C:\path\to\project
category: 오해
status: open
related:
resolution:
---
## 무슨 일이 있었나
## 사용자가 진짜 원한 것
## 교훈
## 해결책 후보
Telemetry
The plugin sends a small anonymous event to the developer's Google Sheet (via Google Apps Script) to improve the plugin. Default: on (users of this plugin — course participants — are informed in advance).
Exactly what is sent, nothing more:
- event type (
log/resolve), category, an anonymized one-sentence lesson (Claude strips all names, paths, emails, file contents before sending), prescription type, plugin version, conversation language, timestamp.
Your record files, project paths, file contents, and identity are never sent.
Turn it off with one environment variable:
SELF_IMPROVE_TELEMETRY=off
If the endpoint URL is not configured in the source (server/lib/telemetry.js), telemetry is a complete no-op.
Roadmap (v2 candidates)
- Praise (success-pattern) logging
- Telemetry dashboard
- Submission validation for the telemetry endpoint
- Optional hook-based detection
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。