Agent skills (skillpacks)
Portable, markdown-defined capabilities that an LLM agent loads — domain expertise, conventions, and step-by-step methods that fire when relevant, often bundled with helper scripts and commands.
Bridge concept (split 2026-06-01). The agent-tooling instances and the open standard (agentskills-spec) now live in the sibling agentic-tooling-wiki; that spoke is the home for the tooling. This page stays in research-wiki because agent-skills is the seam between agent tooling and this wiki’s knowledge-management thesis (see below). Links below resolve cross-wiki.
The A/B seam (why this stays here)
The shared pattern is: capability as markdown content, tool-agnostic, version-controlled, discovered/routed by the agent at runtime — exposed via model-context-protocol or a tool’s native skill directory, and increasingly via the open agentskills-spec. This rhymes with the llm-wiki insight that the substrate is just markdown an LLM reads and acts on: agent skills encode procedures, the LLM wiki encodes knowledge — two faces of one bet. Progressive disclosure (load skill metadata, then body, then resources on demand) is even the procedural analog of the llm-wiki’s index → page drill-down. This is the cleanest reason the two wikis are one super-thesis (“LLMs operating over file-based markdown”) rather than two.
Skill-shaped things that live here (cluster A)
- gbrain — ships 43 curated markdown skills as a single skillpack the installer drops into the agent workspace (and is itself a knowledge base — the A/B straddle).
- llm-wiki-agent — is itself “a coding agent skill” (
/wiki-ingest,/wiki-query, …).
Instances & the standard (now in agentic-tooling-wiki)
agentskills-spec (the open standard) · adk · agentic-seo-skill · claude-financial-services · claude-skills-ppc · agentsys · agent-kanban · gstack · compound-engineering-plugin.
Related
llm-wiki · gbrain · llm-wiki-agent · compound-engineering · model-context-protocol · agentskills-spec