skilld

Generate AI agent skills from your NPM dependencies.

Why?

When using new packages or migrating to new versions, agents often struggle to use the appropriate best practices. This is because agents have knowledge cutoffs and predict based on existing patterns.

Methods of getting the right context to your agent require either manual curation, author opt-in, external servers or vendor lock-in. See the landscape for more details.

Skilld generates agent skills from the references maintainers already create: docs, release notes and GitHub issues. With these we can create version-aware, local-first, and optimized skills.

Made possible by my Sponsor Program 💖 Follow me @harlan_zw 🐦 • Join Discord for help

Features

• 🌍 Any Source: Opt-in - Any NPM dependency or GitHub source, docs auto-resolved
• 📦 Bleeding Edge Context - Latest issues, discussions, and releases. Always use the latest best practices and avoid deprecated patterns.
• 📚 Opt-in LLM Sections - Enhance skills with LLM-generated Best Practices, API Changes, or your own custom prompts
• 🧩 No Agent Required - Export prompts and run them in any LLM (ChatGPT, Claude web, API). No CLI agent dependency.
• 🔍 Semantic Search - Query indexed docs across all skills via retriv embeddings
• 🧠 Context-Aware - Follows Claude Code skill best practices: SKILL.md stays under 500 lines, references are separate files the agent discovers on-demand — not inlined into context
• 🎯 Safe & Versioned - Prompt injection sanitization, version-aware caching, auto-updates on new releases
• 🤝 Ecosystem - Compatible with npx skills and skills-npm

Quick Start

Run skilld in a project to generate skills for your dependencies through a simple interactive wizard:

npx -y skilld

Requires Node 22.6.0 or higher.

Or add a specific package directly:

npx -y skilld add vue

If you need to re-configure skilld, just run npx -y skilld config to update your agent, model, or preferences.

No agent CLI? No problem — choose "No agent" when prompted. You get a base skill immediately, plus portable prompts you can run in any LLM:

npx -y skilld add vue
# Choose "No agent" → base skill + prompts exported
# Paste prompts into ChatGPT/Claude web, save outputs, then:
npx -y skilld assemble

Tips

• Be selective - Only add skills for packages your agent struggles with. Not every dependency needs one.
• LLM is optional - Skills work without any LLM, but enhancing with one makes them significantly better.
• Multi-agent - Run skilld install --agent gemini-cli to sync skills to another agent. The doc cache is shared.

Installation

Global

Install globally to use skilld across all projects without npx:

npm install -g skilld
# or
pnpm add -g skilld

Then run skilld in any project directory.

Per-Project

If you'd like to install skilld and track the lock file references, add it as a dev dependency:

npm install -D skilld
# or
yarn add -D skilld
# or
pnpm add -D skilld

Automatic Updates

Add to package.json to keep skills fresh on install:

{
  "scripts": {
    "prepare": "skilld update -b"
  }
}

FAQ

Why don't the skills run?

Try this in your project/user prompt:

Before modifying code, evaluate each installed skill against the current task.
For each skill, determine YES/NO relevance and invoke all YES skills before proceeding.

How is this different from Context7?

Context7 is an MCP that fetches raw doc chunks at query time. You get different results each prompt, no curation, and it requires their server. Skilld is local-first: it generates a SKILL.md that lives in your project, tied to your actual package versions. No MCP dependency, no per-prompt latency, and it goes further with LLM-enhanced sections, prompt injection sanitization, and semantic search.

Will I be prompt injected?

Skilld pulls issues from GitHub which could be abused for potential prompt injection.

Skilld treats all data as untrusted, running in permissioned environments and using best practices to avoid injections. However, always be cautious when using skills from untrusted sources.

Do skills update when my deps update?

Yes. Run skilld update to regenerate outdated skills, or add skilld update -b to your prepare script and they regenerate in the background whenever you install packages.

CLI Usage

# Interactive mode - auto-discover from package.json
skilld

# Add skills for specific package(s)
skilld add vue nuxt pinia

# Update outdated skills
skilld update
skilld update tailwindcss

# Search docs across installed skills
skilld search "useFetch options" -p nuxt

# Target a specific agent
skilld add react --agent cursor

# Install globally to ~/.claude/skills
skilld add zod --global

# Skip prompts
skilld add drizzle-orm --yes

# Check skill info
skilld info

# List installed skills
skilld list
skilld list --json

# Manage settings
skilld config

Commands

Command	Description
skilld	Interactive wizard (first run) or status menu (existing skills)
skilld add <pkg...>	Add skills for package(s), space or comma-separated
skilld update [pkg]	Update outdated skills (all or specific)
skilld search <query>	Search indexed docs (-p to filter by package)
skilld list	List installed skills (--json for machine-readable output)
skilld info	Show skill info and config
skilld config	Configure agent, model, preferences
skilld install	Restore references from lockfile
skilld remove	Remove installed skills
skilld uninstall	Remove all skilld data
skilld cache	Cache management (clean expired LLM cache entries)
skilld eject <pkg>	Eject skill as portable directory (no symlinks)
skilld assemble [dir]	Merge LLM output files back into SKILL.md (auto-discovers)

Works Without an Agent CLI

No Claude, Gemini, or Codex CLI? Choose "No agent" when prompted. You get a base skill immediately, plus portable prompts you can run in any LLM to enhance it:

skilld add vue
# Choose "No agent" → installs to .claude/skills/vue-skilld/

# What you get:
#   SKILL.md           ← base skill (works immediately)
#   PROMPT_*.md        ← prompts to enhance it with any LLM
#   references/        ← docs, issues, releases as real files

# Run each PROMPT_*.md in ChatGPT/Claude web/any LLM
# Save outputs as _BEST_PRACTICES.md, _API_CHANGES.md, then:
skilld assemble

skilld assemble auto-discovers skills with pending output files. skilld update re-exports prompts for outdated packages.

Eject

Export a skill as a portable, self-contained directory for sharing via git repos:

skilld eject vue                    # Default skill directory
skilld eject vue --name vue         # Custom directory name
skilld eject vue --out ./skills/    # Custom path
skilld eject vue --from 2025-07-01  # Only recent releases/issues

Share via skilld add owner/repo — consumers get fully functional skills with no LLM cost.

CLI Options

Option	Alias	Default	Description
--global	-g	false	Install globally to ~/<agent>/skills
--agent	-a	auto-detect	Target specific agent (claude-code, cursor, etc.)
--yes	-y	false	Skip prompts, use defaults
--force	-f	false	Ignore all caches, re-fetch docs and regenerate
--model	-m	config default	LLM model for skill generation (sonnet, haiku, opus, etc.)
--name	-n		Custom skill directory name (eject only)
--out	-o		Output directory path override (eject only)
--from			Collect releases/issues/discussions from this date (YYYY-MM-DD, eject only)
--debug		false	Save raw LLM output to logs/ for each section

The Landscape

Several approaches exist for steering agent knowledge. Each fills a different niche:

Approach	Versioned	Curated	No Opt-in	Local	Any LLM
Manual rules	✗	✓	✓	✓	✓
llms.txt	~	✗	✗	✗	✓
MCP servers	✓	✗	✗	✗	✗
skills.sh	✗	~	✓	✗	✗
skills-npm	✓	✓	✗	✓	✗
skilld	✓	✓	✓	✓	✓

Versioned — tied to your installed package version. Curated — distilled best practices, not raw docs. No Opt-in — works without the package author doing anything. Local — runs on your machine, no external service dependency. Any LLM — works with any LLM, not just agent CLIs.

• Manual rules (CLAUDE.md, .cursorrules): full control, but you need to already know the best practices and maintain them across every dep.
• llms.txt: standard convention for exposing docs to LLMs, but it's full docs not curated guidance and requires author adoption.
• MCP servers: live, version-aware responses, but adds per-request latency and the maintainer has to build and maintain a server.
• skills.sh: easy skill sharing with a growing ecosystem, but community-sourced without version-awareness or author oversight.
• skills-npm: the ideal end-state: zero-token skills shipped by the package author, but requires every maintainer to opt in.
• skilld: generates version-aware skills from existing docs, changelogs, issues, and discussions. Works for any package without author opt-in.

Telemetry

Skilld sends anonymous install events to skills.sh so skills can be discovered and ranked. No personal information is collected.

Telemetry is automatically disabled in CI environments.

To opt out, set either environment variable:

DISABLE_TELEMETRY=1
DO_NOT_TRACK=1

Related

• skills-npm - Convention for shipping agent skills in npm packages
• mdream - HTML to Markdown converter
• retriv - Vector search with sqlite-vec

License

Licensed under the MIT license.