菜单
Spawn an appropriate subagent to handle a task.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | delegate |
| description | Spawn an appropriate subagent to handle a task. |
| disable-model-invocation | true |
| argument-hint | <task description> |
Delegate the task to a subagent. Do not execute work directly.
/delegate as the task descriptionRoute to the subagent that best matches the task. The roles below are descriptions, not agent names — use your platform's actual agent names.
| Task Type | Subagent Role |
|---|---|
| Find code/files, trace deps | Exploration / investigation |
| Design approach, architecture | Planning |
| Commands, multi-step work, refactor | General-purpose / workhorse |
| Code review * | Review specialist |
| Web research * | Research specialist |
* Fall back to your platform's general-purpose / workhorse agent if no specialist is available.
A fire-and-forget subagent is fine for read-only work (find, trace, review, research) — you consume the result at the end. For editing or iterative tasks (refactor, multi-step implementation), a subagent the user can't interrupt ploughs ahead while objections pile up, leaving a large diff to unwind.
For those, prefer a primitive the user can talk to mid-flight — a
named teammate (below) if the harness supports it, otherwise keep
the work in the main thread rather than a sealed subagent. See
shared/SUBAGENT-STEERABILITY.md
for the full rule.
For editing or iterative tasks, spawn the delegate as a named teammate the user can chat with directly:
delegate-{short-slug} (a 2-3 word
kebab-case summary of the task, e.g. delegate-fix-auth).docs/**/*.md or **/README.md
when behavior, public APIs, or usage patterns change.See shared/AGENT-TEAMS.md for
enablement instructions.
Write an implementation plan to docs/plans/. ALWAYS use this skill — never hand-roll a plan by mimicking files in docs/. Use when the user wants to create a project/implementation plan, when a plan discussed in chat should be persisted, or says /project-plan. Guarantees a spec exists, invoking to-spec if absent. Decomposes into vertical slices with Blocked-by edges, points to the spec for acceptance criteria, and extracts ADRs via to-adr.
Parse test failure output and diagnose root causes in a read-only background subagent, then fix the failures interactively in the main thread. Use when the user shares test output or says "tests fail". Default output path is /tmp/output.
Surface systemic patterns from an investigation as codified conventions or anti-patterns.
Architect a change from idea to actionable artifacts — bootstrap project instructions, research deeply, write a spec, then an implementation plan, each delegated to its skill. Use when the user wants to research a topic, explore a repo, write a spec, create a project plan, or says /architect.
Research a topic or repository deeply and produce a reference document under `docs/research/`. Handles two modes: code research (repo by URL, `org/repo`, or bare name — e.g. "check the spotless repo", "look at github.com/fastly/spotless") and topic research (concepts, technologies, patterns). Use when the user wants to research something, explore a repo, or says /research.
Elicit the user's intent before starting work. Use when a request is vague, when kicking off a new task, when another skill hits a vague request and needs to clarify it, or when the user says "help me with this", "I need something", "let's work on...", "draft a task", or /task.