SkillsHooksPromptsAgentsPersonasModelsPoliciesToolsTemplatesBundlesCategoriesStart here

start here

Two readers, one catalog.

Everything below is true for both: an atom is one JSON file, addressed by class/slug, validated against its class schema, and served verbatim at /atoms/<class>/<slug>.json. What differs is how you get there.

For humans

Inspect, copy, install

Find it. Open a class from the top bar, filter by category, source, or license in the left rail, or press / anywhere to search all 1170 atoms by name, id, description, or tag.
Read it. The page renders the atom's full text as markdown and lists what it references. The footer says who wrote it, where it came from, and its license — license unknown means the source stated none.
Copy it. The rail on the right copies or downloads the operative text: a skill as SKILL.md for .claude/skills/<slug>/, a hook as its script, a prompt as its content, a persona as a rendered identity. Every page also copies the raw JSON.
Install it by id. With the ai CLI on your machine:
skillai skills install commit
hook, wired into every AI clientai hooks install branch-guard
model, from its providerollama pull llama3.2
Publish your own. The builder turns a form into a valid atom for any class. Save it under atoms/<class>/ and open a pull request; the build validates it and checks every reference.

For AI agents

Discover, query, resolve

Discover. One document lists the classes, per-class and per-category counts, every atom id, and the endpoint and schema URLs.
GET https://ai-atoms.com/ai/index.json
GET https://ai-atoms.com/llms.txt
Query. A category returns ids grouped by class. The whole catalog has every atom inline for local filtering.
GET https://ai-atoms.com/categories/{category}.json
GET https://ai-atoms.com/exports/catalog.json
Resolve. Fetch one atom by id; the slug is the id without its class prefix.
GET https://ai-atoms.com/atoms/{class}/{slug}.json
Use the operative field per class.
  • skill → inject system_prompt_fragment; invocation lists slash forms
  • hook → install script; fire on event and trigger; respect blocking
  • prompt → inject content into applicable_turns; composites are pre-resolved
  • persona → render role, voice, tone, constraints, knowledge_boundaries
  • agent → resolve persona, prompts, skills, tools, policies, hooks by id
  • model → call a providers[] entry by model_id, or run its pull_command
Respect provenance. Read provenance.license before redistributing; unknown means the source stated none. The full contract is in /ai/instructions.md.

The six classes

What a runtime does with each one.

Sk Skills · 681Bounded capability with a system-prompt fragment. Drop into .claude/skills/.Hk Hooks · 18Fires on a runtime event. Script ships inline; blocking or advisory.Pr Prompts · 53Typed text injected into the context window: constraint, format, refusal, persona.Ag Agents · 3A persona bound to the prompts, skills, tools, policies, and hooks it runs with.Pe Personas · 47A portable identity: role, voice, tone, work contract, constraints, boundaries.Mo Models · 283Reference data about a model and the providers a runtime can get it from.Po Policies · 54A rule a runtime permits, forbids, or bounds an agent by: boundary, capability, isolation.To Tools · 20An executable affordance: the function the model sees and the side effects to gate.Te Templates · 10A document skeleton a runtime fills: ADR, runbook, handoff, plan, PR, postmortem.Bu Bundles · 1A multi-file skill: agent personas, contracts, templates, docs, and runtime code shipped together.