create-workflow is the authoring counterpart to the rest of the pack: describe
what you want in plain English, and it clarifies, provisions, designs, scaffolds,
verifies, and documents a real, runnable workflow instead of you hand-writing a
.tsx graph and its prompts (see Stages below).
bunx smithers-orchestrator approve RUN_ID --node approve-design, or run
autonomously with --input '{"review":false}'.
When NOT to use it
The clarify stage right-sizes every request before anything is built. A most-trivial edit is routed back to be done directly, and a clear, well-scoped task one strong agent can finish in one context window is routed tosmithers oneshot; both end the run at status
routed-simple with the exact command to run instead. Structure is a cost:
the OrchBench benchmark (benchmarks/orchbench/RESULTS.md in the repo)
measured a solo frontier agent outscoring a three-model review panel at half
the wall clock, so create-workflow only proceeds when the task genuinely
needs ordered stages, durability, approvals, loops, or reuse. To force a
workflow anyway, say so in the prompt.
Stages
- clarify: right-size the request (direct edit,
smithers oneshot, or a real workflow), then turn the freeform ask into a structured, buildable spec (goal, trigger, inputs, stages, loops, human gates, success criteria). - provision: pull the relevant
llms-*.txtdocs, find the closestexamples/template, pick components, and install the worker skills the new workflow needs (bunx smithers-orchestrator skills add). - design: produce the concrete graph blueprint (schemas, tasks, JSX shape, prompts to author).
- approve-design: durable human approval gate (skippable via
review: false). - scaffold: write the workflow
.tsxand its.mdxprompts to.smithers/. - verify:
bunx smithers-orchestrator graphthe new workflow (load + render without executing); a fix agent edits and retries until it renders cleanly. - document: write a skill doc so agents know how to run the new workflow.
Inputs
Monitor it
.smithers/ pack without learning the full
authoring API: it grounds its design in real Smithers components and verifies the
result before handing it back.