Lazy Skill Loader

On-demand skill loading to reduce routing ambiguity, not just context bloat.

Agent runtimes may use progressive disclosure, where the model initially sees only each installed skill's name, description, and path. That still creates semantic overload when many always-installed skills have overlapping descriptions. lazy-skill keeps the always-visible skill surface small and lets the user explicitly summon heavier or niche skills only when needed.

The canonical lazy skill library is ~/.agents/lazy-skills/. ~/.claude/lazy-skills/ is a legacy fallback only.

Index

Skills available for lazy loading (name: sparse keywords - short description):

taste-skill [collection]: taste, frontend, imagegen - Anti-generic frontend design, image-to-code, and premium UI direction
emil-design-eng: emil, design, animation - Emil Kowalski UI polish, component craft, and animation judgment
shadcn-ui [collection]: shadcn, components, registry - Official shadcn/ui component and registry workflow
next-skills [collection]: nextjs, rsc, cache - Next.js best practices, upgrades, and Cache Components
vercel-agent-skills [collection]: react, vercel, performance - Vercel React/Next/RN performance and composition guidance
callstack-react-native-skills [collection]: react-native, callstack, mobile - React Native performance, upgrades, brownfield, and CI artifacts
expo-skills [collection]: expo, eas, mobile - Expo SDK, EAS, dev-client, deployment, modules, and native UI workflows
wshobson-agents [collection]: patterns, architecture, mobile - Broad engineering pattern library including React Native design and architecture
threejs-skills [collection]: threejs, 3d, webgl, graphics - "Three.js skills for 3D graphics (10 skills)"
astro: astro, ssr, islands - Astro project and framework guidance
cloudflare-skills [collection]: cloudflare, workers, pages - Official Cloudflare Workers, Pages, storage, Wrangler, and platform guidance
cloudflare-opennext: opennext, cloudflare, pages - Deploy OpenNext applications on Cloudflare
payload-skills [collection]: payload, cms, migrations - Payload CMS development and migration guidance
remotion-skills [collection]: remotion, video, react - Remotion video creation best practices
inference-video-skills [collection]: video, generation, storyboard - Curated video generation, storyboard, ad spec, and render workflows
mattpocock-skills [collection]: prd, tdd, diagnose - Matt Pocock's engineering, productivity, writing, and workflow skills
bun-runtime: bun, runtime, test - Bun runtime, package manager, bundler, and test runner guidance
hono: hono, api, middleware - Hono API, routing, middleware, testing, and Cloudflare-style app guidance
golang-skills [collection]: go, golang, backend - Go style, testing, concurrency, performance, security, and production patterns
rust-skills [collection]: rust, cargo, systems - Rust ownership, errors, concurrency, unsafe review, ecosystem, and domain patterns
architecture-wise-tree [collection]: architecture, ddd, system-design - Durable architecture, DDD, ADR, refactoring, and system-design judgment
quality-constitution: quality, mutation, complexity - Agent-operable quality gates for duplication, CRAP risk, mutation, and review queues

Keep this index sparse. It is a card catalog, not mini documentation. Avoid long trigger lists, exhaustive "use when" rules, or overlapping descriptions that recreate the same routing noise this skill is meant to avoid.

Behavior

On Invocation

If $ARGUMENTS provided, filter index to matching keywords
If blank, show full index
Always ask user which skill to load before reading - never read all skills
Use Read tool to load the selected skill file

Path Resolution

Resolve the lazy skills root in this order:

$AGENTS_LAZY_SKILLS_DIR, if set
~/.agents/lazy-skills
~/.claude/lazy-skills (legacy fallback)

Single skills - try in order:

<lazy-root>/<name>.md
<lazy-root>/<name>/SKILL.md

Collections (marked with [collection] in index):

Read <lazy-root>/<name>/README.md to show available skills
Ask user which specific skill to load
Resolve the selected skill by either:
- folder path: <lazy-root>/<name>/skills/<skill-name>/SKILL.md
- frontmatter name: recursively search <lazy-root>/<name>/**/SKILL.md for name: <skill-name>
Read the resolved SKILL.md

After Loading

Skill content is now in context
Follow the loaded skill's instructions
No separate "invoke" step - it's just knowledge now

Adding Skills to Index

Edit this file's Index section:

- **skillname**: keyword1, keyword2, keyword3 - "One-line description"

Then place the skill file at either:

~/.agents/lazy-skills/skillname.md (single file)
~/.agents/lazy-skills/skillname/SKILL.md (full skill folder, e.g., cloned repo)
or a custom root via $AGENTS_LAZY_SKILLS_DIR

Examples

Single skill:

User: /lazy-skill stripe

Claude: Found 1 matching skill:
- **stripe**: payments, billing, checkout - "Stripe API operations"

Want me to load it?

User: yes

Claude: [Reads ~/.agents/lazy-skills/stripe.md]
Loaded stripe skill. What would you like to do?

Collection:

User: /lazy-skill threejs

Claude: Found 1 matching collection:
- **threejs-skills** [collection]: threejs, 3d, webgl - "Three.js skills (10 skills)"

Want me to show the available skills in this collection?

User: yes

Claude: [Reads ~/.agents/lazy-skills/threejs-skills/README.md]

Available skills in threejs-skills:
| Skill | Description |
|-------|-------------|
| threejs-fundamentals | Scene setup, cameras, renderer |
| threejs-geometry | Shapes, BufferGeometry, instancing |
| threejs-materials | PBR, shader materials |
... (7 more)

Which one should I load?

User: fundamentals

Claude: [Reads ~/.agents/lazy-skills/threejs-skills/skills/threejs-fundamentals/SKILL.md]
Loaded threejs-fundamentals. Ready to help with Three.js scene setup.

<instructions>$ARGUMENTS</instructions>