SkillLoop — Self-Healing AI Skills

The Problem

Your team built a library of AI skills — prompt templates for code review, doc generation, SQL analysis, and dozens of other tasks. Everyone uses the same skills. Consistency is high. Life is good.

Then reality kicks in:

• A skill works for common cases, but edge cases trip it up. Users correct the AI mid-conversation to get the right output.
• Each person solves the same problem independently — in their own chat history, invisible to everyone else.
• There is no path to feed those individual fixes back into the shared skill.
• The skill stays broken. New team members hit the same wall. Veteran users waste time re-explaining the same fix.

The result? A growing gap between how your skills should work and how they actually work — widening silently with every conversation.

The Solution

SkillLoop watches your AI conversations transparently, detects when you correct a skill, extracts the issue, and — when enough people hit the same problem — generates a ready-to-review Pull Request that fixes the skill. You just keep working. The skills get better on their own.

Core Value

Key Features

1. Zero-Intrusion Detection

SkillLoop uses each AI platform’s native hook system — no skill code changes, no wrappers, no middleware. Install the adapter once, and every conversation is observed transparently.

2. Two-Path Analysis

When a correction is detected, SkillLoop prefers local analysis — spawning a background AI agent (Claude Code or Copilot CLI) on your machine. If no local CLI is available, it falls back to the server. Either way, you get the same result.

3. Smart Clustering

Similar issues from different users are grouped automatically using sentence embeddings. Duplicate PRs for the same underlying problem are prevented — the skill owner sees one clear PR, not a flood of noise.

4. Automatic PR Generation

When a cluster reaches the evidence threshold, SkillLoop fetches the current skill file, generates an improved version, and opens a Draft PR with the problem pattern, root cause, proposed diff, and evidence count.

5. Web Admin Dashboard

A built-in admin page for managing skill registries, server settings, correction patterns, and monitoring signals — no config files to edit inside containers.

Quick Start

1. Start the server — docker-compose up -d (one command, one time)
2. Install the adapter — pip install skillloop-client && skillloop-client install copilot --url http://your-server:8080
3. Use your AI skills as usual — no changes to your workflow
4. SkillLoop detects corrections — when you say “no, that’s wrong” or “try again”, it captures the signal
5. Review the PR — when enough evidence accumulates, a Draft PR appears in your skill repo

Download

The client package has zero external dependencies — pure Python stdlib.

Supported AI Platforms

Privacy & Data

• All data stays on your infrastructure — self-hosted server, local database, no cloud dependency
• No account required — no sign-up, no SaaS, no subscription
• Session data is never stored on disk by the client — streamed to the server on session end
• LLM calls stay internal — configure your own LLM endpoint (DashScope, OpenAI-compatible, or local model)

Roadmap

SkillLoop is currently in prototype stage.

What’s available now:

• Zero-intrusion hooks for Claude Code, Copilot CLI, and VS Code Copilot Chat
• Two-path architecture (local agent + server fallback)
• Correction detection with configurable regex patterns (Chinese + English)
• Issue extraction and clustering via LLM + embeddings
• Web dashboard and admin UI
• Multi-repo skill registry with scan-from-repo

Coming soon:

• Codex CLI, Cursor, and Cline adapters (Phase 2)
• Automatic PR generation with evidence-based diffs
• Auto-merge mode for high-confidence fixes
• GitLab and Gitea provider support

Community & Support

• Report an issue
• Read the documentation
• View the source code