菜单
Set up a new project on Swarm — create its channel and seed it with intelligence records. Use when connecting a repository to Swarm for the first time, or when the user asks to onboard, initialize, or set up a project on Swarm.
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
Checkpoint the current conversation back to its Swarm channel. Use ONLY when the user explicitly asks to checkpoint, share, or save their work to Swarm — never invoke automatically, because sharing is gated on the channel's capture mode and the user's intent.
Use when working in a project connected to Swarm, a shared context bus over MCP. At the start of a task, load shared context with get_context_for_injection; at meaningful checkpoints, share back per the channel's capture mode. Triggers whenever the Swarm MCP tools (get_context_for_injection, capture_context, share_session, list_channels) are available.
Turn working context into durable channel canon — promote or create records, link them, and supersede replaced ones. Use when the user asks to curate, record, promote, or save findings as Swarm records, or to relate or supersede existing records.
Load shared Swarm context for the active channel. Use at the start of a task in a Swarm-connected project, or whenever the user asks to load or refresh Swarm context.
Work the produce-loop on a Swarm Project. Use when the user asks you to work on their Swarm project, address or review the change requests (recommendations) on a Swarm Project, or produce the changes a human has recommended — you author the edits to the Project's Artifacts and open a review gate for a human to approve. Scoped to a Swarm Project (channel + slug), not any local project folder.
Hand off curated context to another person or agent with a Swarm handoff bundle. Use when the user asks to hand off, transfer, or package context for a teammate or successor, or to inspect or acknowledge a handoff bundle.
| name | swarm-onboard |
| description | Set up a new project on Swarm — create its channel and seed it with intelligence records. Use when connecting a repository to Swarm for the first time, or when the user asks to onboard, initialize, or set up a project on Swarm. |
Onboarding creates ONE new channel and seeds it from the repository in front of you. It never reads from, adopts, or inherits the workspace of another channel.
The onboarding ritual is:
list_channels only to guard against an exact duplicate: a channel whose key is the slug of this channel's name in the same target workspace. A similar or shorter name (e.g. acme vs acme-billing) is a different project — ignore it. Never call get_context_for_injection on another channel during onboarding, and never inherit its workspace.onboard_project tool to create the channel and seed it with those records.<!-- swarm-channel: <workspace-slug>/<channel-key> --> line to the project's own instructions file (CLAUDE.md, AGENTS.md, a .cursor/rules/*.mdc file — whichever this client uses; create one if none exists). This is a hard step, not optional.If you are unsure about the channel name or the workspace, ask the user before calling onboard_project rather than assuming and continuing.
If your client surfaces MCP prompts, the onboard-project prompt walks this same ritual interactively and is the canonical, server-rendered source — prefer it when it is available. Some clients do not expose MCP prompts; when the prompt is not available, follow the steps above directly (they call the same onboard_project tool).
After onboarding, treat the channel as the project's shared memory: load it at the start of each task with get_context_for_injection, and checkpoint at the end per the channel's capture mode.