The manager’s job.
Give the system a defined task, a review boundary, and a result you can measure.
“Help with operations” is a wish. “Prepare a quote from this request, show the source, and wait for review” is the beginning of a job. The difference is work the manager has to do.
Write a useful job description for a first AI workflow.
Start with a task, not a personality
Naming the agent and giving it a friendly tone does not tell it what a finished job looks like. Start by watching the person who does the work now. Where does the request arrive? What do they open? Which exceptions make them stop?
Do not write the ideal process from memory. Ask someone to walk through a recent example, including the awkward parts. Those details often explain why a tidy demo does not survive a normal working day.
Write the job in five lines
- Input: what arrives, where it arrives, and what must be present.
- Output: the thing someone can open and use.
- Rules: the sources and constraints that govern the work.
- Review: when to stop, and who decides what happens next.
- Measure: how you will compare the work before and after.
Add one more instruction: show what is missing or uncertain. A visible gap is easier to resolve than an invented answer hidden inside a polished document.
Make the example concrete
Example job: preparing a freight quote
- Input
- A quote request with a lane and pickup date.
- Output
- A draft in the quote sheet, linked to the rate source.
- Rules
- Use the current approved rate card and customer terms.
- Review
- Flag an unknown lane. Wait for a pricing lead before sending.
- Measure
- Time to a reviewed quote, including correction and review time.
Replace every line with the reality of your own workflow.
Choose how much judgment the task needs
Anthropic distinguishes predefined workflows from agents that choose their own next steps. That is a useful design distinction—not a ranking of sophistication. Start with a fixed sequence when you already know the steps. Add model-directed decisions where the task actually requires them.
For the quote example, receiving a request, checking required fields, and creating a draft may have a predictable order. Resolving an unfamiliar exception may need judgment. You can keep the predictable parts explicit and send the exception to a person.
Measure the whole job
A draft that appears instantly but takes longer to correct has not necessarily improved the process. Count preparation, review, correction, and follow-up. Compare similar cases, and keep unusual cases visible rather than quietly removing them from the sample.
Track whether the intended users actually use the system. If they avoid it, ask why. Missing context, inconvenient placement, and unclear responsibility are problems to solve with them—not reasons to blame the team.
Try it with your work.
Fill in the five-line job description for one workflow. Ask the person doing it today to mark anything that is wrong, missing, or too vague.
Use the workflow worksheetCheck your understanding
The draft takes less time, but the reviewer spends longer correcting it. Did the workflow improve?
Show the answer
Not enough information yet. Compare total time and quality across the whole job, including review and rework. Faster generation alone is not the outcome.
Sources & further reading
Anthropic: Building effective agents The distinction between a predefined workflow and model-directed action.
Source pages checked 19 October 2026. Our examples and recommendations are editorial interpretation, not findings from those publications.Ontara Academy · Written lesson · 2026-10-19.