using-skills

name	using-skills
description	Meta skill that guides the agent on when to pick an existing skill, when to create a new one, and when to install an MCP server. Adapted from obra/superpowers' "using-superpowers" and addyosmani's "meta" skill for personal-AI use.
version	0.1.0
tags	["meta","skills"]
metadata	{"hermes":{"created_by":"user"},"soulclaw":{"category":"meta","triggers":["有哪些技能","你会什么","用哪个技能","怎么做这个","which skill","what can you do"],"capabilities":["skill_routing"],"related_skills":["skill-authoring","mcp-discovery","skill-vetter"]}}

using-skills

Trigger

Pick this skill when the user asks what you can do, which skill applies to their task, or when the right move is unclear. Also activate internally before starting any non-trivial task so the agent does a one-pass "do I already have a skill for this?" check.

Principle

Use the right skill for the task; install a new capability only when none of the installed options fit.

Core capabilities are always available but hidden from the user view. User-installed extensions surface through /list.

Decision Tree

1. Does an installed skill match the intent?

   • Yes → route the turn through that skill. Done.
   • No → step 2.

2. Can the task be done with existing tools (read_url, web_search, code_execution, read_file, etc.) without a new skill?

   • Yes → run it directly, no new skill needed.
   • No → step 3.

3. Is the workflow durable (will it repeat across sessions / days)?

   • Yes → propose to author a new skill via the skill-authoring skill. Confirm with the user before creating.
   • No → one-off execution with existing tools.

4. Does the task need an external service the agent cannot reach with current tools (e.g. GitHub API, Notion, a video download service)?

   • Yes → route through the mcp-discovery skill: search the MCP ecosystem, propose ONE candidate, install after confirmation.
   • No → the task is probably out of scope; say so plainly.

Inputs

• user_intent: a short summary of what the user is asking for.
• current_inventory: the harness inventory (what skills / MCP / plugins are installed). Fetch via /api/harness/extensions or by calling tool_search when browsing built-in tools.

Steps

1. Summarize user intent in one line.
2. Walk the decision tree above.
3. If creating or installing, confirm with the user first — never install without a yes/no.
4. Record the decision briefly if the path is non-obvious ("chose skill X because Y").

Verification

• The routed / suggested skill actually matches the intent.
• No new skill is created when an existing one would fit.
• No MCP is installed without user confirmation.
• For skill creation, run skill-authoring rather than hand-writing frontmatter.

Failure Signals

• The same one-off task spawns a new skill (skill inflation).
• A skill matches by name but not by workflow, causing wrong results.
• The user says "you already have that" — signals a missed inventory check.
• Multiple MCP installs proposed in one turn — too aggressive.

Output Format

When the user asks "what can you do", reply with a compact 3-5 line summary grounded in the live inventory, not a hardcoded list. Mention core capabilities exist without enumerating them:

当前可用：<N> 个已安装的用户技能 / <M> 个 MCP 服务器 / <K> 个插件。
常见场景：<一句话举例>。
核心能力（内置隐藏）：旅行、天气、写作、调试、规划、研究等。
需要新功能？告诉我做什么，我去找合适的 skill 或 MCP。