---

name: run-iflow description: 向助手发出协助请求，适用中华文化理解、中文古文理解、中国特色作品创作、小红书与抖音的宣发文案编写、编写测试用例、信息检索等任务场景，同时也是其他专家助手的强力替补与协助，可以作为大部分认为的备选候补方案。

功能简介

通过 MCP 工具向一位外部助手发出协助请求，请对方在一个上下文干净的、运行时隔离的、独立环境中，执行中华文化理解、中文古文理解、中国特色作品创作、小红书与抖音的宣发文案编写、编写测试用例、信息检索等任务。

使用方法

使用 mcp__plugin_headless-knight_runCLI__iflow 工具来向助手发出请求，可传入参数包括：

• prompt: string，从上下文整理出的完整的任务描述，必填参数
• systemPrompt: string，从上下文整理出的需要该助手遵守的系统提示词，用于约束他的行为，可选
• workDir: string，工作目录，默认为当前目录，可选
• model: string，指定使用哪个模型，取值为"GLM-4.6"、"Qwen3-Coder-Plus"、"DeepSeek-V3.2"、"MiniMax-M2"或"Kimi-K2-0905"等模型代号，可选，默认为 MiniMax-M2
• env: object，自定义环境变量，键值对，可选

模型选择

• GLM-4.6: 擅长中文内容创作、中文市场分析、中国文化相关的教育与问答等任务
• Qwen3-Coder-Plus: 擅长阿里云用户或需要进行模型精调和私有化部署等任务
• DeepSeek-V3.2: 擅长与编程相关的任务，如代码生成、Bug 修复、代码解释、算法设计、编写单元测试等
• Kimi-K2-0905: 适用于需要处理大量文本信息的场景，如信息检索、长篇内容总结、知识库问答等
• MiniMax-M2: 适用于上述各项任务

相关环境变量

如果不另外设置，则自动使用当前对话上下文中的配置

• IFLOW_API_KEY: Codex CLI 的 API_Key
• HTTP_PROXY: HTTP 代理地址
• HTTPS_PROXY: HTTPS 代理地址
• ALL_PROXY: 默认代理地址
• IFLOW_CLI_COMMAND: Codex 命令行地址或名称（确保在 PATH 中包含了路径）
• NODE_ENV: 环境参数，比如 "development"

Claude Skills Registry

Core repo: logic + index + site.
Main repo (merged artifact): https://github.com/majiayu000/claude-skill-registry
Data repo (skills archive): https://github.com/majiayu000/claude-skill-registry-data
Authority: core workflows are canonical; main is a publish mirror.

The most comprehensive Claude Code skills registry — updated daily with the latest skills

What is this?

The largest searchable index of Claude Code skills, aggregated from GitHub and community sources.

Three ways to use:

1. Web Search - Fast browser-based search
2. sk CLI - Terminal package manager
3. API - Direct JSON access

Repo layout note: core owns workflows/pipeline logic, data stores skills/**, and main is generated from core + data. See SCHEME2_SPLIT.md.

Generated Mirror Notice

This repository is the merged publish artifact for browsing and compatibility consumers. Source changes for discovery, download, security scanning, index generation, search, Pages, or publish orchestration belong in claude-skill-registry-core. Archived skill body and archive metadata issues belong in claude-skill-registry-data.

• Do not submit normal source PRs here unless the change is explicitly main-owned, such as publish workflow routing or mirror-only repository metadata.
• Release status: this mirror does not cut independent GitHub Releases for daily generated data. Use provenance/merge-source.json, registry_summary.json, and the generated timestamps to identify the exact core_sha and data_sha behind an artifact.
• Use issues here only for mirror artifact problems, such as mismatched generated files, broken compatibility paths, or incorrect provenance after publish.

Highlights

• Massive Skill Index - Deduplicated, high-quality registry (see badge for live count)
• Rich Categories - Development, Testing, DevOps, Design, and more
• Daily Updates - Automated crawling/validation by core scheduled workflows
• Quality Indexed - Metadata, descriptions, and star counts
• Lightweight Search - Gzip-compressed index for fast client-side search

Operational Ownership

• Core: source of truth for workflows, crawling, scanning, and index/site generation
• Data: canonical archived skill tree (skills/**)
• Main: publish artifact for merged browsing/compatibility consumers
• Publish contract: core dispatches main publish with pinned core_sha + data_sha
• Non-goal for main: main does not initiate canonical syncs, archive updates, or index generation

Quick Start

Option 1: Web Search

Visit https://majiayu000.github.io/claude-skill-registry-core/

For clone/update tips on large repositories, see docs/FAST_CLONE.md.

Option 2: CLI (sk)

# Install sk
go install github.com/majiayu000/caude-skill-manager@latest

# Search skills
sk search testing
sk search pdf
sk search --popular

# Install a skill
sk install anthropics/skills/skills/docx

Option 3: Direct API

The public JSON surface is versioned in docs/artifact-api-contract.md. Use that contract when integrating with pointer files, manifests, shards, gzip variants, and deprecation markers.

# Startup search index and bounded full-search shards
curl https://majiayu000.github.io/claude-skill-registry-core/search-index.json
curl https://majiayu000.github.io/claude-skill-registry-core/search-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry-core/search-shards/part-000.json

# Deduplicated catalog index with quality/security/install signals
curl https://majiayu000.github.io/claude-skill-registry-core/search-index-lite.json
curl https://majiayu000.github.io/claude-skill-registry-core/quality-index.json
curl https://majiayu000.github.io/claude-skill-registry-core/quality-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry-core/quality-shards/part-000.json
curl https://majiayu000.github.io/claude-skill-registry-core/security-index.json
curl https://majiayu000.github.io/claude-skill-registry-core/security-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry-core/security-shards/part-000.json
curl https://majiayu000.github.io/claude-skill-registry-core/ranking-index.json
curl https://majiayu000.github.io/claude-skill-registry-core/ranking-index-manifest.json
curl https://majiayu000.github.io/claude-skill-registry-core/ranking-shards/part-000.json

# Lightweight registry summary (counts only)
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry-core/main/registry_summary.json

# Full registry manifest and shards (merged publish artifact)
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry/main/registry-manifest.json
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry/main/registry-shards/00.json

# Compatibility registry pointer (merged publish artifact)
curl https://raw.githubusercontent.com/majiayu000/claude-skill-registry/main/registry.json

# Category manifest and bounded parts
curl https://majiayu000.github.io/claude-skill-registry-core/categories/index.json
curl https://majiayu000.github.io/claude-skill-registry-core/categories/development/manifest.json
curl https://majiayu000.github.io/claude-skill-registry-core/categories/development/part-000.json

# Legacy category pointer
curl https://majiayu000.github.io/claude-skill-registry-core/categories/development.json

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  Layer 1: Data Collection                                       │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐          │
│  │ GitHub Crawl │→ │ Download     │→ │ Security     │          │
│  │ (discover)   │  │ (sync)       │  │ (scanner)    │          │
│  └──────────────┘  └──────────────┘  └──────────────┘          │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Layer 2: Index Generation                                      │
│  ┌────────────────┐  ┌────────────────┐  ┌────────────────┐    │
│  │ search-index   │  │ categories/    │  │ featured.json  │    │
│  │ .json          │  │ *.json         │  │ (featured set) │    │
│  └────────────────┘  └────────────────┘  └────────────────┘    │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Layer 3: Consumption                                           │
│  ┌────────────────┐  ┌────────────────┐  ┌────────────────┐    │
│  │ Web UI         │  │ sk CLI         │  │ API            │    │
│  │ (GitHub Pages) │  │ (Go)           │  │ (JSON)         │    │
│  └────────────────┘  └────────────────┘  └────────────────┘    │
└─────────────────────────────────────────────────────────────────┘

Search Index Schema

// Lightweight index for fast search
interface SearchIndex {
  v: string;           // Version (date)
  t: number;           // Indexed skill count (always s.length)
  s: SkillMini[];      // Skills array
}

interface SkillMini {
  n: string;           // name
  d: string;           // description (truncated 80 chars)
  c: string;           // category code (dev, ops, sec, etc.)
  g: string[];         // tags (max 5)
  r: number;           // stars
  i: string;           // install path
}

---

Directory Structure (Core)

claude-skill-registry-core/
├── registry.json           # Lightweight core registry metadata
├── docs/                   # GitHub Pages
│   ├── index.html          # Web search UI
│   ├── search-index-lite.json
│   ├── search-index.json   # Compatibility pointer to bounded shards
│   ├── search-index-manifest.json
│   ├── search-shards/      # Bounded full search index parts
│   ├── featured.json       # Featured skills snapshot
│   └── categories/         # Category manifests and bounded parts
├── sources/                # Data sources
│   ├── anthropic.json
│   ├── community.json
│   └── skillsmp.json
├── scripts/                # Build scripts
│   ├── build_search_index.py
│   ├── discover_by_topic.py
│   ├── security_scanner.py
│   └── ...
└── (no committed skills/)  # skills/** lives in registry-data; mounted in CI when needed

The merged claude-skill-registry publish artifact additionally contains registry-manifest.json and registry-shards/**.

---

Categories

Canonical category slugs, short codes, governance status, and heuristic keywords live in taxonomy/categories.yaml. Pipeline scripts read that file instead of keeping their own category lists. Legacy category names are diagnostic inputs only; they are routed to review and are not accepted as publish categories. The canonical category proposal is documented in docs/plan/canonical-taxonomy-proposal.md, and the current target count report is docs/plan/canonical-category-target-counts.json. Validate the taxonomy itself with python scripts/check_taxonomy_governance.py. To validate publish output against canonical categories, pass repeated --publish-category <slug> values. To review category quality across the archive before moving anything, run python scripts/audit_category_quality.py --skills-dir skills. The default audit uses metadata and paths for a fast full-archive pass; add --include-frontmatter when checking frontmatter/category drift. The audit also reports non-standard nested skill paths such as category/category/skill/SKILL.md. For those layout issues, use python scripts/normalize_skill_depth.py --skills-dir skills --json to review the exact move plan before applying it. For semantic reclassification, generate a review-only migration plan with python scripts/plan_category_migration.py --skills-dir skills --output category-migration-plan.json. The plan records action, confidence, source categories, target category, keyword signals, and reason for every proposed change; it does not move files. For a bounded second-pass model review, set MIMO_API_KEY and run python scripts/review_category_plan_with_llm.py --plan category-migration-plan.json --output category-llm-review.json --checkpoint-jsonl category-llm-review.checkpoint.jsonl --resume. The default endpoint is https://token-plan-sgp.xiaomimimo.com/v1 with mimo-v2.5-pro. The default request uses --thinking disabled and --max-completion-tokens 1024 so the model output budget is reserved for the required JSON; raise the token value for larger reviews, or pass --thinking default when using a non-MiMo endpoint that does not accept the provider-specific thinking field. The report remains review-only: it records model category, confidence, decision, parse status, token/thinking policy, and evidence without modifying the archive. The checkpoint file is append-only JSONL, so interrupted long reviews can resume without re-calling the model for completed candidates.

Category counts are published in categories/index.json; full category payloads are available through categories/<category>/manifest.json and bounded part-*.json files. The legacy categories/<category>.json URL is now a small compatibility pointer. The publish path runs python scripts/check_category_artifacts.py --categories-dir docs/categories so legacy category files cannot silently grow back into large full-payload JSON. Common category codes include:

Category	Code	Description
development	dev	Development tools, frameworks
data	dat	Data processing, analysis
design	des	UI/UX design, frontend
testing	tst	Testing, QA, automation
devops	ops	DevOps, CI/CD, infrastructure
documents	doc	Document creation (docx, pdf)
productivity	pro	Productivity and automation
product	prd	Product management
security	sec	Security, auditing
marketing	mkt	Marketing, content, SEO

---

Roadmap

Current Status

• Index count tracked by the badge (core registry.json)
• Archive size: tracked by badge (raw SKILL.md count from stats.json)
• Daily auto-update via GitHub Actions
• Security scanning for all skills
• sk CLI for installation

In Progress

• Lightweight search index (gzip-compressed; see stats.json)
• Web search UI (GitHub Pages)
• GitHub Pages deployment (https://majiayu000.github.io/claude-skill-registry-core/)

Planned

• AI semantic search (vector similarity)
• Skill recommendations (based on usage)
• Version tracking for skills
• Skill quality scoring
• API rate limiting and caching

---

Contributing

Add Your Skill

Option 1: Submit via Issue

1. Open an issue
2. Use the "Add Skill" template
3. Provide: repo URL, name, description, category

Option 2: Submit via PR

1. Fork majiayu000/claude-skill-registry-core
2. Add your skill to sources/community.json:

{
  "name": "your-skill-name",
  "repo": "your-username/your-repo",
  "path": "optional/path/to/skill",
  "description": "What your skill does",
  "category": "development",
  "tags": ["testing", "automation"]
}

3. Submit a PR to majiayu000/claude-skill-registry-core (not claude-skill-registry, which is a generated publish artifact)

Report Issues

We welcome feedback! Please open an issue for:

• Bugs - Search not working, incorrect data
• Feature requests - New categories, better search
• UX improvements - Web UI, CLI enhancements
• Data quality - Duplicate skills, wrong categories

👉 Open an Issue

Contribute Code

# Clone the core repo (authoritative pipeline repo)
git clone --filter=blob:none --sparse https://github.com/majiayu000/claude-skill-registry-core.git
cd claude-skill-registry-core

# Pull only what you need (add more paths later as needed)
git sparse-checkout set --cone docs scripts sources schema

# Install dependencies
uv venv
source .venv/bin/activate
uv pip install -r requirements.txt

# Build search index locally
python scripts/build_search_index.py --registry registry.json --output docs

# Test web UI
cd docs && python -m http.server 8000
# Visit http://localhost:8000

See docs/FAST_CLONE.md for more options (existing clones, getting full checkout, Windows notes).

---

The Agent Infra Stack

This project is one layer of an open-source stack for running coding agents (Claude Code, Codex) as serious infrastructure. Every piece works standalone; together they close the loop:

In this stack, the registry is the Extend layer — the front door where you discover skills. Scan third-party skills with argus before installing, and keep agents honest at runtime with vibeguard.

Layer	Project	What it does
Extend	claude-skill-registry ◀ you are here	Discover and search community Claude Code skills
Extend	spellbook	Cross-runtime skills for Claude Code, Codex, and multi-agent workflows
Trust	argus	Static install-time scanner for supply-chain attacks (npm / PyPI / crates.io)
Trust	vibeguard	Rules, hooks, and guards against hallucinated or unverified agent changes
Remember	remem	Local-first persistent memory for Claude Code and Codex sessions
Orchestrate	harness	Rust agent orchestration platform — rules, skills, GC, observability
Route	litellm-rs	High-performance Rust AI gateway — 100+ LLM APIs via OpenAI format
Keep	keepline	Session command center — monitor, recover, never lose agent work

---

Related Projects

Project	Description
caude-skill-manager	CLI tool for installing skills (sk)
anthropics/skills	Official Anthropic skills
SkillsMP	Web-based skill marketplace
awesome-claude-skills	Curated skill list

---

License

MIT License applies to the registry code/pipeline only - see LICENSE for details.

Third-Party License & Attribution

Third-party skills under skills/** keep their original licenses and copyright ownership.

• Repository-level MIT does not relicense third-party skill content.
• Every imported skill metadata file should include:
  • author
  • source_url
  • license
  • copyright
  • permission_note
  • distribution (compatible or restricted)
• restricted entries are not MIT-compatible by default and require explicit upstream permission before redistribution/use.
• Notices are generated by compliance checks into THIRD_PARTY_NOTICES.md.
• Metadata compliance runs in advisory mode by default to avoid blocking ingestion; strict blocking can be enabled when needed.

---