ワンクリックで
write-a-skill
Create new agent skills. Use when user wants to create, write, or build a new skill.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Create new agent skills. Use when user wants to create, write, or build a new skill.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Bun is an all-in-one JavaScript/TypeScript toolkit that ships as a single binary. It includes a runtime (drop-in Node.js replacement), package manager, test runner, and bundler.
Use when configuring .oxlintrc.json or lint-related CI
WXT web extension framework — entrypoints, manifest generation, content scripts/UI, storage, messaging, modules, CLI, browser targets, publishing. Use when building, debugging, configuring, migrating, or publishing browser/web extensions with WXT.
Read this skill before serving as orchestrator
Read before writing user-facing docs — READMEs, tutorials, how-to guides, reference docs, or any prose for human readers. Use when creating or substantially editing docs aimed at users rather than agent-facing files like AGENTS.md.
Delegate work to configured subagents with single-agent, chain, parallel, async, forked-context, and intercom-coordinated workflows. Use only agents discovered from user or project Markdown agent files.
| name | write-a-skill |
| description | Create new agent skills. Use when user wants to create, write, or build a new skill. |
A skill is a prompt that gets loaded into a future agent's context when it picks up a relevant task. Your job is to capture two things from the user, then translate them into that prompt:
Everything else is in service of those two.
Do not over-prescribe. The agent reading the skill later is capable and will have full context about its task. You're priming its reasoning, not scripting it. Prefer transmitting intent, examples, and quality bars over step-by-step procedures. If a step is genuinely load-bearing (e.g. a non-obvious command, a required format), state it; otherwise trust the agent.
A good skill reads like a senior colleague briefing you on what matters, not a runbook.
Interview the user enough to actually understand the skill. Ask about real triggers, real examples, and where prior attempts went wrong. If they hand you a rough draft, your job is to sharpen it, not pad it out.
When you're unsure whether to include something: leave it out. The user wants leverage, not coverage.
The description is the only thing a future agent sees when deciding whether to load this skill. It must:
Keep it tight. If the description is vague, the skill will never get loaded at the right moment.
A skill lives at skills/<name>/SKILL.md with frontmatter name and description. If the skill needs supporting reference material that would bloat SKILL.md, split it into sibling files and link to them; otherwise keep everything in one file. Length should match the actual complexity of the domain — some skills are ten lines, some are a few hundred. Don't pad to hit a template.
Re-read the draft as if you were the future agent loading it cold. Ask:
Cut the filler. Then show the user.