メニュー
Use this skill when the user wants to capture the current conversation as a reusable skill. Analyze the session and call create_skill to propose one — the user reviews the full body in the approval card before it's written to disk.
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
| name | propose-skill |
| description | Use this skill when the user wants to capture the current conversation as a reusable skill. Analyze the session and call create_skill to propose one — the user reviews the full body in the approval card before it's written to disk. |
The user wants to capture the work in this conversation as a reusable skill — a procedure the agent can re-run on demand via /<skill-name>. Your job: analyze the session, then call create_skill once with the proposed parameters. The user will see your proposal in the approval card and accept or reject.
/propose-automation instead.refactor-component, audit-secrets)Call create_skill exactly once with:
name: lowercase, hyphens only, must start with a letter, max 48 chars. Example: "refactor-component".description: one line — what the skill does AND when to use it. This is what the agent reads to decide whether to activate the skill, so include keywords that match user intent.body: the full SKILL.md body, after the frontmatter (the system adds frontmatter from name + description). Markdown. Start with a # Title heading. Include step-by-step instructions, gotchas, examples. Write it for the next agent invocation — that agent has no memory of this conversation.Before the tool call, write 2-3 sentences in plain prose explaining:
The user reads your prose first, then sees the structured args in the approval card.
body is for the next invocation. That agent has no memory of this conversation. Be explicit about steps, expected inputs/outputs, and known gotchas.body. The system adds name: and description: from the other args.If the conversation is a one-off Q&A or exploration with no procedure worth codifying, say so plainly in 1-2 sentences and DO NOT call create_skill. Don't manufacture a proposal.
Use this skill when the user wants a memory feed — a scheduled automation that fetches targeted data from their integrations (gmail/calendar/slack/web) and keeps ONE memory page current, updated in place. Examples; "track my PR queue", "keep an invoices page", "week-ahead brief".
Implement a spec end-to-end — recon readers map the codebase, an architect pins every cross-file contract in a blueprint, parallel builders implement disjoint workstreams from it, adversarial review, then a final test gate. Use for multi-file feature work. Run via the workflow tool by name.
Find -> adversarially verify -> rank issues over a target (a diff, dir, file, or glob). Use for code review, a scoped bug hunt, a security or perf pass. Run via the workflow tool by name.
Answer a question about a target by fanning out parallel readers over derived angles, then synthesizing a cited answer. Use to understand a subsystem, research a question, or trace how/where something works. Run via the workflow tool by name.
Decide between approaches — generate N diverse proposals from different lenses, score each against criteria with independent judges, then synthesize a decisive recommendation. Use for design choices, tradeoff calls, "which approach should we take". Run via the workflow tool by name.
Use this skill when the user wants to turn the current conversation into a scheduled automation. Analyze the session and call create_automation to propose one — the user reviews the full prompt + schedule in the approval card before it's saved.