← ClaudeAtlas

flow-workflowslisted

Author and run flow workflows — turn a recurring multi-step process into a saved, replayable workflow.json (typed agent/script/subflow nodes + edges) and execute it with xdog-flow. Agent nodes can call a coding-agent CLI (claude/codex) for their step, so no provider/API key is needed. Use when asked to crystallize/automate/save a process as a workflow, or to build a typed pipeline (classify→route, draft→critique→revise, research→summarize).
litterzhang/x-dog · ★ 1 · AI & Automation · score 71
Install: claude install-skill litterzhang/x-dog
# flow — author & run workflows Use this skill to turn a recurring, multi-step process into a **saved, replayable workflow** and then run it. flow is a typed, single-machine workflow kernel: you describe nodes (agent / script / human / subflow) and the typed edges between them as JSON, then `xdog-flow run` executes it — or `xdog-flow generate` compiles it to a self-contained Python module. The key move: an **agent node can call a coding-agent CLI** (`claude`, `codex`) for its step instead of a hosted provider. So this CLI can crystallize your own workflows and run them, with each agentic step shelling back to a CLI — no API key or provider to configure. ## When to use - The user asks to "save / crystallize / automate this process as a workflow." - A task is a fixed pipeline of steps (classify → route, draft → critique → revise, research → summarize) worth replaying deterministically around the LLM steps. - You need typed, inspectable orchestration rather than an ad-hoc prompt chain. ## Workflow JSON at a glance ```jsonc { "name": "cli-triage", "entry": "classify", // optional; else derived from $in edges "state": { "report": "..." }, // $in seed values (the workflow's inputs) "nodes": [ { "id": "classify", "type": "agent", "backend": "claude-cli", // run this step via the claude CLI (no provider) "model": "sonnet", // CLI --model (alias or id); optional "inputs": [{ "name": "report