skilld-skilld

name	skilld-skilld
description	Generate AI agent skills from npm package documentation. ALWAYS use when writing code importing "skilld". Consult for debugging, best practices, or modifying skilld.
metadata	{"version":"2.0.0","generated_by":"cached","generated_at":"2026-05-16T00:00:00.000Z"}

skilld-dev/skilld skilld@2.0.0

Tags: latest: 2.0.0

References: package.json • README • Docs • Issues • Releases

Search

Use skilld search "query" -p skilld instead of grepping .skilld/ directories. Run skilld search --guide -p skilld for full syntax, filters, and operators.

API Changes

This section documents version-specific API changes — prioritize recent major/minor releases to understand what's new in v2.0.0.

• NEW: skilld prepare — new command (v1.5.0) for restoring references and syncing shipped skills in package.json hooks source

• NEW: skilld author — new command system (v1.5.0) for maintainers to generate and publish skills directly with their packages source

• NEW: skilld list --outdated — flag added (v1.4.0) to filter and display only skills with available version updates source

• NEW: skilld search --filter — flag added (v1.5.0) for filtering search results by type (e.g., issue, discussion, release) using JSON syntax source

• NEW: skilld search --limit — flag added (v1.5.0) to cap the number of search results returned source

• NEW: skilld search --guide — flag added (v1.5.0) to display full search syntax, filters, and operators for advanced queries source

• NEW: skilld cache --stats — flag added (v1.6.0) to display cache statistics including size and age of LLM-generated sections source

• NEW: skilld cache --clean — flag added (v1.6.0) to remove stale and expired cache entries source

• NEW: No agent mode — workflow added (v1.1.0) allowing skilld to run without an agent CLI, generating portable prompts for any LLM instead source

• NEW: AI OAuth providers — authentication flow added (v1.3.0) with UX rework to support multiple LLM provider logins source

• NEW: Private GitHub repo fallback — feature added (v1.2.0) allowing skill docs and sync to resolve from private GitHub repositories source

• NEW: Incremental search index updates — search system improved (v1.3.0) to rebuild only changed docs instead of all-or-nothing rebuilds, reducing sync time source

• NEW: LLM-generated sections — skills now support optional sections for Best Practices, API Changes, and custom prompts curated by LLMs instead of raw docs source

Also changed: Dangling symlink handling · Isolated lockfile cache copies · Skip LLM on full cache hit · jsonc-parser for surgical edits

Best Practices

• Be selective about skill installation — only add skills for packages your agent actually struggles with, not every dependency. Unnecessary skills add context overhead without benefit source

• Use skilld prepare in your package.json prepare script to auto-restore symlinks, auto-install shipped skills, and get notified of outdated packages on every install source

• Add LLM enhancements (Best Practices, API Changes, custom prompts) only when the base skill needs refinement — skillD works effectively without any LLM if docs are clear source

• Use skilld search "query" with the -p package flag for semantic search across docs instead of reading files manually — retriv embeddings surface exact answers faster source

• Organise skills via curated collections when you install multiple related packages — one command pulls the full set aligned with best practices from trusted curators like @harlan-zw or vercel-labs source

• Target specific agents with skilld add npm:package --agent claude-code (or cursor, gemini-cli, etc.) to install skills into the right agent directory without manual setup source

• For no-agent workflows, choose "No agent" when prompted — you get a base SKILL.md immediately, plus PROMPT_*.md files you can paste into ChatGPT/Claude web, then assemble the outputs with skilld author assemble source

• Use --force flag when debugging skill generation or cache issues — it clears all caches and regenerates from scratch, avoiding stale LLM output or partial doc fetches source

• Eject skills as portable, self-contained directories for sharing via git repos using skilld author eject vue --out ./skills/ — consumers then install via skilld add gh:owner/repo with zero LLM cost source

• Ship skills directly in npm packages (skills/ directory) so consumers get them automatically via skilld prepare — no LLM needed on their end and zero opt-in required source

• Use --debug flag to save raw LLM output to logs/ for each SKILL.md section, useful for inspecting generated best practices, API changes, or verifying prompt output before assembly source

• Run skilld install --agent gemini-cli to sync skill references across multiple agents — the doc cache is shared, avoiding duplicate downloads and keeping all agents in sync source

• Follow the skill evaluation pattern: before invoking a skill, check YES/NO relevance to the current task — only invoke skills that actually apply, preventing wasted context on tangential skills source

• Treat all skill content (especially GitHub issues used for enhancements) as untrusted input — skilld runs in permissioned environments and sanitises prompt injection vectors, but always vet skills from untrusted sources source