Agent Skills & Subagents
A skill is a folder with a SKILL.md that an agent loads when the task calls for it. A subagent is a named agent with its own instructions that the main agent can hand work to. Store each one once in LynxPrompt; the download and lynxp pull write it to the folders the agents read.
Agent Skills
A skill blueprint holds one SKILL.md: frontmatter with name and description, then the instructions. The name is 1 to 64 lowercase letters, digits and hyphens, with no hyphen at either end and no two in a row, and it is the folder name. LynxPrompt refuses a name that breaks this, because Copilot, Junie and Cline refuse to load such a skill.
| Written to | When | Read by |
|---|---|---|
.agents/skills/<name>/SKILL.md | Always | Codex, Cursor, Copilot, Gemini CLI, Antigravity, OpenCode, Zed, Junie, Devin, Cline, Kilo, Goose, Amp, Augment, Trae, Factory, Qwen Code, Kimi Code, Mistral Vibe, Pi, Warp, Crush, OpenHands and more |
.claude/skills/<name>/SKILL.md | Always | Claude Code, which reads only this folder |
.kiro/skills/<name>/SKILL.md | When you pick Kiro | Kiro |
.firebender/skills/<name>/SKILL.md | When you pick Firebender | Firebender |
The .claude/skills copy is the skill as you wrote it. The other copies keep only the fields of the Agent Skills format (name, description, license, compatibility, metadata, allowed-tools), so Claude-only fields such as when_to_use, context, paths, hooks or model stay where Claude Code reads them. A blueprint holds the SKILL.md only, not scripts or other files in the skill folder.
Subagents
A subagent blueprint holds one Markdown file: frontmatter with name and description (plus any Claude Code fields such as tools or model), then the agent's instructions.
| Written to | When | Read by |
|---|---|---|
.claude/agents/<name>.md | Always, as written | Claude Code, Cursor, Copilot in VS Code, Goose |
.agents/agents/<name>.md | Always, name and description only | Antigravity, Goose, OpenHands, Kimi Code, Devin CLI |
.github/agents/<name>.agent.md | When you pick Copilot | GitHub Copilot |
.codex/agents/<name>.toml | When you pick Codex | OpenAI Codex (name, description, developer_instructions) |
lynxp push also picks up subagents in .cursor/agents, .gemini/agents, .opencode/agents, .kiro/agents, .junie/agents, .factory/droids, .qwen/agents, .augment/agents, .trae/agents, .kilo/agents, .cline/agents and .devin/agents.
Web and CLI
- On the web, pick Agent Skill or Subagent as the format when you create a blueprint. Its download is a
.zipthat unpacks into the project root with every copy; tick Kiro, Firebender, Copilot or Codex to add theirs. lynxp pullwrites the same copies, adding Kiro, Firebender, Copilot or Codex when the project already has their folder (.kiro,.firebender,.github/agents,.codex). One copy is tracked; the others are plain copies, never symlinks. A copy you already have with different content is kept, with a warning, unless you pass--yes.lynxp pushrecognises a skill or subagent in any of the folders above. With--all, the copies LynxPrompt writes are pushed once, and after a push orlynxp syncthe copies already in the project are rewritten from it. A skill in another tool's own folder, such as.github/skills, is pushed on its own. A subagent stored as Codex TOML or Kiro JSON keeps its single file and path.lynxp convert .claude/agents/reviewer.md codex-agentwrites one copy for another tool. Targets:shared-skill,claude-skill,kiro-skill,firebender-skill,claude-agent,shared-agent,copilot-agent,codex-agent.