Agent Skills Framework
Agent Skills Framework packages repeatable agent know-how as a folder with a required SKILL.md (YAML frontmatter plus markdown instructions). We adopt it as the only realistic way to make agent behavior portable across models and providers.
Summary
| When | Action |
|---|---|
| Building portable ops for coding agents | Use the layout: one folder per skill, SKILL.md with strong name and description |
| One-off prompts in a single tool | Skip; inline rules or tool-native config may be enough |
| Importing community skills | Vet each repo; use Agent Skills - Sources; run in sandbox first |
Portability: Tool-native rules, system prompts, and vendor config do not travel. A SKILL.md folder ports across Claude Code, Cursor, Codex, Gemini, and other runtimes with path changes at most. That is the practical baseline for shared agent behavior in a multi-model estate.
Layout: skills/<skill-name>/SKILL.md plus optional scripts/, references/, assets/. The description field is the primary trigger signal. Keep the body under ~500 lines. Push detail into reference files.
In this vault: agent-skills/skills/ mirrors the pattern. gbrain two-layer pages and identity files in AGENTS.md / CLAUDE.md sit alongside skills.
Details
Required Frontmatter
| |
Skill Types (Informal)
| Type | Purpose | Example |
|---|---|---|
| Capability | Teach a concrete tool or API | apple-reminders, proofread |
| Behavioral | Layer policy on other skills | reminders-gtd, proofread-blog |
| Domain | Your outcome-specific workflow | morning-briefing, company-research |
Runtimes in the Garden
- Claude Code:
.claude/skills/or sharedagent-skills/ - Cursor / cursor-agent:
.cursor/skills/ - Gemini CLI, Codex, Cline, OpenCode: tool-specific paths; core
SKILL.mdoften ports with path changes only
See Agent Skills - Sources for official docs, awesome lists, and marketplaces.