{
 "acronym": "",
 "adoption": {
  "by": null,
  "level": "observed",
  "probe_run": null
 },
 "aliases": [
  "Agent Skills",
  "SKILL.md"
 ],
 "analog": "Package README / plugin manifest",
 "collisions": [
  "MCP \"Skills over MCP\" working group (a related but distinct proposal for discovering skills through MCP itself)"
 ],
 "discovery": {
  "mechanism": "Filesystem/package convention: a directory containing a SKILL.md file with YAML frontmatter (required fields name, description) plus Markdown instructions; agents discover skills by scanning known skill directories, loading only name+description at startup (\"progressive disclosure\") and the full body only on activation",
  "path": "<skill-name>/SKILL.md (directory name must match the frontmatter name field)"
 },
 "dispute": null,
 "evidence": [
  {
   "claim": "Agent Skills is a lightweight open format: a skill is a directory with a required SKILL.md (YAML frontmatter with required name and description fields, optional license/compatibility/metadata/allowed-tools) plus Markdown instructions and optional scripts/, references/, assets/ directories; agents load skills via three-stage progressive disclosure (discovery of name+description, activation loads the full body, execution loads bundled resources on demand). Validation tooling ships as skills-ref.",
   "date": "2026",
   "label": "VERIFIED",
   "url": "https://agentskills.io/specification"
  },
  {
   "claim": "\"The Agent Skills format was originally developed by Anthropic, released as an open standard, and has been adopted by a growing number of agent products. The standard is open to contributions from the broader ecosystem\" — development happens on github.com/agentskills/agentskills and a public Discord. A client showcase lists dozens of named adopting products including Claude Code, Claude.ai, Cursor, GitHub Copilot, VS Code, Gemini CLI, ChatGPT & Codex, JetBrains Junie, Amp, OpenHands, Goose, Letta, Factory, Roo Code, OpenCode, Snowflake Cortex Code, Databricks Genie Code, and Spring AI, among ~35 named logos.",
   "date": "2026",
   "label": "VERIFIED",
   "url": "https://agentskills.io"
  }
 ],
 "form": "document-format",
 "governance": "Originally developed and released as an open format by Anthropic; now described by its own site as \"open to contributions from the broader ecosystem\" with development happening in the community-run github.com/agentskills/agentskills org and a public Discord",
 "group": null,
 "history": [
  {
   "date": "2026-09-05",
   "note": "seeded from Exa Agent research run"
  },
  {
   "date": "2026-09-06",
   "note": "verified against agentskills.io and agentskills.io/specification"
  }
 ],
 "id": "anthropic-agent-skills-format",
 "implementation": "multi-implementation",
 "last_verified": "2026-09-06",
 "name": "Agent Skills",
 "native_status": "Active open format, described on agentskills.io as adopted by a growing number of agent products",
 "origin": {
  "date": "2025",
  "org": "Anthropic"
 },
 "part_of": null,
 "parties": [
  "agent-site"
 ],
 "phase": "living",
 "probe": {
  "method": "filesystem read (not a network probe)",
  "on_absent": "no SKILL.md at that path, or frontmatter missing required fields: not a valid Agent Skill",
  "parse": "YAML frontmatter containing required `name` (max 64 chars, lowercase/digits/hyphens) and `description` (max 1024 chars) fields, per the agentskills.io specification",
  "path": "{skills-dir}/{skill-name}/SKILL.md"
 },
 "question": "How can reusable, on-demand procedural instructions and bundled resources be packaged so any compatible agent can load and run them?",
 "scope": "coding-tools",
 "spec_url": "https://agentskills.io/specification",
 "step": "act",
 "summary_md": "Agent Skills is an open packaging format for reusable procedural knowledge: a \"skill\" is a directory containing a required `SKILL.md` file (YAML frontmatter with a `name` and `description`, plus optional `license`, `compatibility`, `metadata`, and experimental `allowed-tools` fields) followed by Markdown instructions, and optionally `scripts/`, `references/`, and `assets/` subdirectories for executable code, longer documentation, and templates respectively [1]. It solves the problem of giving an agent domain-specific, team- or user-specific know-how — a legal review checklist, a data pipeline procedure — as a portable, version-controllable unit that works across any compatible agent product, rather than being locked into one vendor's prompt-engineering conventions [2].\n\nThe format's core mechanism is progressive disclosure in three stages: at startup an agent loads only the `name` and `description` (roughly 100 tokens) for every available skill; when a task matches a description, the agent loads the full `SKILL.md` body (recommended under 5,000 tokens); only when the instructions call for it does the agent load bundled scripts, references, or assets [1]. This keeps large skill libraries cheap to keep \"on hand\" without bloating context.\n\nAgent Skills does not define how an agent discovers which skill directories exist in the first place (that is left to each host's own conventions — e.g. a fixed `.claude/skills/` or `~/.codex/skills/` directory) and does not define a runtime protocol for invoking a skill over a network; it is purely a file-format and loading-behavior convention, distinct from the related but separate \"Skills over MCP\" working-group effort that is exploring discovering skills *through* MCP itself, which this row does not cover (cross-reference — see mcp record).\n\nAs of this verification (September 2026), the format was \"originally developed by Anthropic\" but its own site now states it is \"released as an open standard\" and \"open to contributions from the broader ecosystem,\" with development happening in a dedicated `agentskills/agentskills` GitHub organization and public Discord rather than solely inside Anthropic's repos [2] — a materially more open governance picture than the seed record's plain \"Anthropic\" / \"Agent Skills ecosystem\" labels captured. A client showcase names roughly 35 adopting products spanning coding agents (Claude Code, Cursor, GitHub Copilot, VS Code, JetBrains Junie, Amp, OpenHands, Goose, Roo Code, OpenCode, Factory), general assistants (Claude.ai, Gemini CLI, ChatGPT & Codex), and enterprise platforms (Snowflake Cortex Code, Databricks Genie Code, Spring AI) [2]. This is direct, first-party observation of a broad adopter list rather than a systematic count, so adoption here is recorded as `observed`, not `measured`. What remains open: no version number is attached to the specification itself (it is presented as a living document), and the relationship/overlap between this format and the MCP \"Skills over MCP\" working group's discovery mechanism has not been reconciled by either project as of this check.",
 "track": "community",
 "version": {
  "date": null,
  "label": "unversioned living spec (agentskills.io/specification)"
 }
}