leehow/pipiui/create-subagent
- community
Checking that the SKILL.md file is still in the repository.
Create a least-privilege schema v1 PipiUI subagent package through the bundled subagent_manage tool.
Install
npx vskill@latest install leehow/pipiui/create-subagentSKILL.md
Create a PipiUI subagent
Agent, Skill, and Prompt
Use this only in the main session. This Skill is guidance for the current main session; it is not an Agent and grants no tools. Dispatched workers run skill-free and must not load this Skill. A Prompt is only task text, not a reusable definition or permission grant.
The bundled subagent_manage tool is the authority for agent package scaffolding, validation, permission summaries, and installation. Bundled Skill names are retained and bundled-first: create-subagent is the built-in workflow for this task, so a same-named user Skill cannot replace it. Do not copy a schema by hand or write AGENT.md directly when that tool is available. The template beside this Skill is documentation/test material only; runtime rules come from the bundled subagent extension's agents.ts parser.
Collect only what is needed
Ask only for information absent from the request:
- Scope —
userorproject. - Name — concise lower-kebab-case name.
- Description — one clear sentence about role and outcome.
- Mode / optional model —
read-onlyorworker; omit model unless the
user needs one.
- Minimum capabilities — filesystem, shell, web, exact MCP tool names,
desktop, and delegation.
Do not ask for accounts, login, credentials, marketplaces, remote installs, or infrastructure details. The least-privilege baseline is read-only with no filesystem/shell/web/MCP/desktop/delegation, worktree: none, and a report.
Tool-first workflow
- If the user wants to inspect existing roles, call
subagent_manage({action:"list", scope:"user"|"project"|"both"}) or inspect. Read its diagnostics; a shadowed or malformed definition is not a reason to silently invent a replacement.
- Call `subagent_manage({action:"scaffold", scope, name, description, mode?,
model?, capabilities?, worktree?, deliverable?, tools?, prompt?})`. It returns schema v1 content, the exact target path, parser diagnostics, and the capability/tools permission summary.
- Call
subagent_manage({action:"validate", scope, name, draft})on the
proposed content. Use the returned structured errors/warnings and permissionSummary; do not imitate the parser in prose.
- Before any install, show the user the proposed scope/path, mode/model,
worktree/deliverable, declared capabilities, explicit tools narrowing, and the tool's runtime caveats. In particular: global disabled-tools and missing extensions can still remove tools; desktop: requestable needs the existing global Computer Use host gate plus a per-task grant; delegation is depth and runtime-role limited.
- After the user approves that preview, call
subagent_manage({action:"install", scope, name, draft, overwrite?}). Omit overwrite by default. Set overwrite:true only when the user explicitly approved replacing that exact existing package. Report the tool's final install path and verification result.
install only writes a validated standard package under the selected user or project agent root. A project install writes <project>/.pi/agents/<name>/AGENT.md; it still requires the existing runtime project-agent confirmation when that role is run. It rejects bundled targets, traversal, symlink escape, and arbitrary output paths.
Capability choices
read-onlymust useworktree: none; it cannot use
filesystem: workspace-write.
- Use
worktree: isolatedonly for a worker that actually needs isolation. - MCP is false by default. If needed, give only exact registered
mcp_<server>_<tool> names. Never use true, *, a server-wide grant, or a guessed name.
desktop: requestableis only a request for per-task desktop access. Never
put computer or open_application in tools.
web: trueis retrieval capability, not a browser grant. Never put
browser in tools.
- Set delegation only for a real orchestration role. It does not let a worker
load Skills or self-grant desktop/search access.
toolsis optional and only narrows capability-derived tools; it cannot add
a tool or bypass a capability.
Keep the flow local, direct, and preview-first. No remote market, account, or pairing ceremony is needed.
Security
- Security scanPassed, 90 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingElectron/resources/runtime/built-in-skills/create-subagent/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/leehow/pipiui/create-subagent)