Skip to main content
init seeds two things side-by-side: a .tsx orchestration skeleton that handles execution, and an .mdx prompt file you edit to change what the agent does. For most workflows you never touch the .tsx file.

​
The two-file model

FileWhat it isWho edits it
.smithers/workflows/hello.tsxDurable run, agent wiring, output schemaRarely - only when you need branching, schemas, or loops
.smithers/prompts/hello.mdxThe agent’s instruction - what you actually changeYou, always
The .tsx file imports and renders the .mdx file at runtime. Editing the prompt changes agent behavior on the next run - no build step, no TypeScript.

​
Your first edit

After bunx smithers-orchestrator init, open .smithers/prompts/hello.mdx: 
Write a short, friendly one-sentence greeting for {props.name}.

This is your hello-world workflow. Edit this file to change what the agent does -
it is plain Markdown, no TypeScript required. Anything in {curly braces} is a
value passed in from the workflow (here, props.name).
Change the instruction - for example, ask for a haiku: 
Write a haiku about {props.name}.
Run it again: 
bunx smithers-orchestrator workflow run hello --name Ada
The agent now writes haiku. The .tsx file didn’t change.

​
Template syntax

Props come from the workflow input and are interpolated with {props.key}: 
Summarize this pull request for a non-technical stakeholder:

> {props.title}

Keep it under two sentences. Plain language only.
Any input field the workflow declares is available as props.<field> in the prompt.

​
Create a new MDX-backed workflow

Use the built-in create-workflow workflow: 
bunx smithers-orchestrator workflow run create-workflow
Or scaffold by hand:
  1. Create .smithers/prompts/my-task.mdx - write the instruction.
  2. Copy .smithers/workflows/hello.tsx, rename the workflow name and prompt import.
  3. Run bunx smithers-orchestrator workflow run my-task.

​
When to graduate to TypeScript

Stay in MDX as long as possible. Add TypeScript only when you need:
  • Multi-step branching - different tasks depending on a prior output
  • Parallel fan-out - multiple tasks running simultaneously
  • Validated structured output - Zod schemas on task results
  • Loops - retry until a score threshold passes
See the TypeScript SDK for all of that. Start with MDX, graduate only when the task demands it.