跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

import-skill

Import a Rome skill from a URL, a pasted SKILL.md, or a plain-language description — validate it against Rome's frontmatter parser, preview the final file, and install it into the user-skills carrier app after the guardian confirms.

rome-os · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add rome-os/rome --skill import-skill

技能说明


name: import-skill description: Import a Rome skill from a URL, a pasted SKILL.md, or a plain-language description — validate it against Rome's frontmatter parser, preview the final file, and install it into the user-skills carrier app after the guardian confirms. tools: [Read, Write, Edit, Bash, Glob, Grep, WebFetch]

Import a skill

Turn what the guardian brings you into something installed — a URL to an existing skill, a pasted SKILL.md, or a description of one they wish existed. Brainstorm with them: propose, listen, revise. Don't install anything without getting a clear go-ahead.

Most of what you're handed is a plain skill for the user-skills carrier. The involved ones — a setup step, real interaction, a multi-step process — are better off as their own app. Sorting which is which, with the guardian, is part of the job: see "Skill, or a dedicated app?" before you install.

Classify the input

  • URL → see "Resolving a URL".
  • Pasted SKILL.md (frontmatter + body, or close to it) → validate, preview.
  • Description (prose, a wish, a workflow explained) → see "Generating".
  • Empty or ambiguous → ask one short question. Don't guess.

Resolving a URL

  • github.com/<owner>/<repo>/blob/<ref>/<path> → fetch the raw form https://raw.githubusercontent.com/<owner>/<repo>/<ref>/<path>.
  • github.com/<owner>/<repo>/tree/<ref>/<dir> → list it via https://api.github.com/repos/<owner>/<repo>/contents/<dir>?ref=<ref> and download the SKILL.md plus its auxiliary files.
  • Bare repo → look for SKILL.md at the root, then under skills/.
  • Anything else → fetch it; if it's skill-shaped, proceed; if it's prose describing a workflow, offer to generate a skill from it.

If the source holds more than one skill (several SKILL.md under skills/, a repo of them, a page linking many), don't pick for the guardian: list each one you found — path, name, and description — and ask which to import. They may choose one, some, or all; import each as its own pass through the rest of this flow (validate → preview → install), so every skill gets its own confirmation.

On a network error or 404, say so and ask for a corrected link.

Generating from a description

The guardian is the domain expert; you're the author. Ask at most 2–3 focused questions, only where the description is genuinely open:

  • Trigger — when should an agent reach for this? (Becomes the description line, which is also the search index.)
  • Procedure — the concrete steps, commands, decision points. "Check the logs" isn't an instruction; "run X and look for Y" is.
  • Tools — which built-ins the steps need.
  • Boundaries — what it should refuse or hand off.

Draft the SKILL.md yourself and preview it. The body is second-person, imperative, concrete — instructions for a future agent, not human docs.

Skill, or a dedicated app? Decide before you install

A skill is pure instructions an agent reads. Some requests are apps wearing a SKILL.md, and the carrier gives them a worse version of what they asked for. Once you understand the request, judge it. App-shaped signals — any one means it wants to be its own app:

  • an init / setup step before it's usable (credentials, first-run config);
  • it needs to interact — ask and react, not just run top-to-bottom;
  • a real multi-step process, its own data / operations / UI, or background / scheduled execution.

A shell command or two inside otherwise-pure instructions is still a skill. When the signals are real, put the choice to the guardian — don't pick for them:

  • Import as a skill → the user-skills carrier (the flow below).
  • Build a dedicated app → a brand-new app of its own, friendly and interactive.

Only continue into validate/preview/install when they choose the skill.

Building a dedicated app

Read the coding:app_creation skill with read_skill and follow it to scaffold, build, and install a brand-new app — its OWN app id, never user-skills or the carrier. Carry what you've learned from the guardian as the brief: what the thing does, its init/setup step, the interactions it needs.

Design a friendly flow: an init step that gets the guardian set up, interaction that guides rather than dumps questions, and — where it fits — a conversational agent they talk to, with the app's entry dropping straight into that chat. (A pure inputs → steps → result automation isn't an app — coding:app_creation will redirect it to coding:workflow_creation, as it should.)

Validate against Rome's parser

Rome parses frontmatter with regexes, not YAML. A violating skill installs but is silently dropped from the catalog. Enforce on every skill:

  1. Starts with a --- line and has a closing ---.
  2. name: is a single lowercase kebab-case line matching the install folder.
  3. description: is a single line — collapse folded (>) or multi-line ones.
  4. tools: is only read inline (tools: [A, B, C]) — convert dash-lists; drop one you can't convert and say so.
  5. Other keys are ignored, harmless — keep them.

Auto-fix 2–4 and note each fix in the preview. Then check list_skills for a name collision (propose an alternative rather than shadow), and make sure the description carries the trigger words someone would search — search_skills matches name/description/tools, never the body.

Preview, then confirm

Show the guardian the complete final SKILL.md in a fenced block (and any auxiliary files), a note of every fix you made, and the install path user-skills/skills/<name>/. Ask, and wait. Anything short of a clear yes is feedback — revise and preview again.

Install into the carrier

Skills install into one carrier app, user-skills, at <custom app authoring directory from your Runtime Context>/user-skills (~/.rome/<profile>/projects/apps/user-skills).

If it doesn't exist, create it — a minimal hand-rolled layout, no build toolchain, so the coding:app_creation scaffold doesn't apply. It

阅读完整来源 ↗