메뉴
Use when you need to dispatch one or more fast, cheap subagents for parallel or sequential work — research, summarize, transform, verify, batch-process, or any task that doesn't require a specialized agent. Invoke via the Agent tool with subagent_type "atelier:minion".
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when the user runs "/insights-audit", asks to "validate the insights report", "audit the report against my projects", "check what the report got wrong", or wants to cross-check an /insights-generated HTML report against actual GitHub repos and local project directories.
This skill should be used when the user asks to "end session", "capture state", "session summary", "what changed this session", "summarize repos", "write session notes", or wants to capture multi-repo state at the end of a work session.
Fill in the Patterns & Surprises section of the latest session reflect file and write the completed review to .ctx/logs/self-review/. Use when the user runs "/self-review", asks to "fill in the reflect", "write the session review", or "what patterns did we hit".
Use when working on the rslm project — Recursive Language Model inference engine in Rust. Triggers on "rslm", "RLM", "recursive language model", "rhai built-ins", "ctx_grep", "ctx_slice", "rlm_call", "rslm-core", "rslm-providers", "rslm-store", or any question about running queries, ingesting documents, or writing Rhai scripts for the RLM loop.
This skill should be used when the user asks to "review this diff", "ai review", "security review my changes", "review before push", "check my code for issues", or wants an LLM-based security and correctness review of uncommitted or unpushed changes.
This skill should be used when the user asks to "generate commit message", "write a commit msg", "ai commit", "generate a conventional commit", "write my commit", or wants an LLM to draft a commit message from the staged diff.
| name | minion |
| description | Use when you need to dispatch one or more fast, cheap subagents for parallel or sequential work — research, summarize, transform, verify, batch-process, or any task that doesn't require a specialized agent. Invoke via the Agent tool with subagent_type "atelier:minion". |
| model | haiku |
| agent | minion |
| effort | low |
| argument-hint | [task description] |
| allowed-tools | ["Agent(minion)","Read","Bash"] |
atelier:minion is a fast haiku-powered agent that adapts to any task. Use this skill
to decide when and how to dispatch minions, how many to run in parallel, and how to
synthesize their results.
| Situation | Use minion? |
|---|---|
| Quick read + summarize of 1–5 files | Yes |
| Batch task across N independent items | Yes (one per item) |
| Verify a condition (file exists, test passes, etc.) | Yes |
| Transform or reformat content | Yes |
| Complex multi-step implementation | No — use forge |
| Code review | No — use sentinel or oxidizer |
| Handoff / session management | No — use handoff/handon |
Agent(
subagent_type: "atelier:minion",
prompt: "<clear, self-contained task description>"
)
The prompt must be fully self-contained — minion has no prior context. Include:
Send multiple Agent tool calls in a single message. Cap at 5 concurrent.
Agent(subagent_type: "atelier:minion", prompt: "Summarize /path/to/file-a.md")
Agent(subagent_type: "atelier:minion", prompt: "Summarize /path/to/file-b.md")
Agent(subagent_type: "atelier:minion", prompt: "Summarize /path/to/file-c.md")
Task: <one sentence>
Input: <file paths, commands, or data>
Output: <exactly what to return — format, length, structure>
Constraints:
- <any tool restrictions, read-only, etc.>
After minions return, synthesize in the parent context — don't dispatch another minion to summarize minion output unless the synthesis is itself a large batch job.