Designing a multi-step workflow
What you'll be able to do
- Decompose a big task into discrete steps and route each step to the right tool
- Pick a stronger or cheaper model per step instead of one model for everything
- Recognize that chained focused calls beat one giant prompt, and map a real workflow with checkpoints
One mega-prompt rarely wins
It is tempting to cram the whole job into one enormous prompt. It almost never works well. The more you stuff in, the more your real instructions get buried, and the harder it is to tell which part of the output went wrong.
Decompose
Break the task into discrete steps, each with a single purpose: gather, then draft, then check, then format. When each step does one thing, you can inspect its output and fix the weak link without rerunning everything.
Route
Send each step to the tool that does it best. A grounded lookup goes to NotebookLM. Careful drafting goes to Claude. A quick transform can go anywhere. Routing is just matching the step to the tool instead of forcing one tool to do all of it.
Model per step
Heavy reasoning steps deserve a strong model. A trivial reformat does not; a cheaper, faster one is plenty. You do not have to run the flagship on every step, and not doing so saves time and money where it does not matter.
Pass the baton cleanly
One step's output becomes the next step's input, so define the handoff. Be explicit about what shape leaves one step and enters the next. A sloppy handoff is where multi-step workflows quietly fall apart.
Checkpoints between steps
Put a human or an evaluation gate wherever an error would compound: before money moves, before something sends, before anything gets filed. A mistake caught at step two is cheap. The same mistake caught at step six is expensive.
Draw it before you build it
Sketch the workflow as boxes and arrows before you start running prompts. A quick map beats discovering the shape of the workflow mid-task, halfway through, when you realize two steps were in the wrong order.