A one-page workflow brief with an owner, a useful output, an evidence bar, and a stop condition.
Begin with the moment that repeats
Choose a piece of work you have already done more than once. A weekly research brief, a support handoff, or a draft product update is easier to understand than a request to transform an entire business. Describe the trigger and the person waiting for the result.
Write down the current steps before choosing an agent. Which steps follow fixed rules? Which require interpreting language? Which involve a judgment you cannot yet explain? A spreadsheet, script, or checklist may already solve the fixed part. Reserve a model for the part that benefits from it.
Make the result observable
An instruction such as ‘research the market’ leaves the endpoint to the model. Specify the object you want to inspect: a table of source-backed claims, a draft with unanswered questions, or a comparison with explicit trade-offs.
Choose the evidence standard before generating the answer. For a research brief, each material claim might need a dated primary source and an explicit uncertainty note. These criteria structure the exercise. The sources themselves still need scrutiny.
Trigger: Friday planning meeting Owner: Product lead Input: Three public product announcements Output: One-page comparison with dated source links AI may: Extract claims and draft the comparison Human decides: Which claims affect the roadmap Stop when: A material claim has no supporting source
Give the agent a small useful job
Separate preparing an answer from acting on it. ‘Draft a reply’ and ‘send a reply’ are different permissions. If the output can change another person's situation, name the person who can approve that change and the exact evidence they need to see.
Start with one role. Add another only when you can name the independent responsibility it contributes. A reviewer who repeats the same prompt on the same information may add little. The complexity of the team should follow observed needs.
Try the brief against a difficult input
Use a synthetic or public example first. Remove a source, insert conflicting dates, or make the requested action exceed the role. Does your brief tell the agent to stop, ask, or expose the uncertainty? If the failure has no defined outcome, revise the brief.
Save your initial brief and the revision. The difference between them is useful learning material: it shows the assumption you had to make explicit. A brief becomes portable when another person can understand both its purpose and its limits.
Put the idea to work
Try it yourself
Choose a weekly task. In Mission Studio, write its trigger, owner, input, output, evidence standard, and stop condition. Then invent one missing-source input and revise the brief so it handles that case explicitly.
Inspect your result
- Another person can identify the promised output.
- Fixed rules and model judgment have separate jobs.
- Preparation does not silently authorize an external action.
- A missing or contradictory source has a defined response.
These are reflection prompts, not an assessment or certification. Use public or synthetic inputs.
Write your brief in Mission Studio ↗Sources and context
Background on simple, composable workflows. The article is dated December 2024 and is not a current API specification.