Turn an expert process into a repeatable AI workflow

Pick a job with a recognizable finish

Use a deliverable you can inspect: a research brief, a draft reply, an experiment plan, or a reconciliation checklist. Avoid beginning with a whole profession. A narrow job makes it easier to identify which facts change the answer and what a successful handoff looks like.

Write the decisions you normally make silently

Explain why you reject one option and choose another. For an interview-plan review, you might flag leading questions, separate facts from assumptions, and require each question to connect to a research objective. Those decisions are the method. A polished introduction about expertise is not a substitute.

Design a missing-input path

Real users rarely bring perfect information. Tell the AI which omissions prevent progress and which assumptions it may state explicitly. If no research goal is supplied, ask for one. If an example interviewee is missing, use a labeled fictional example instead of inventing a real customer.

For each conclusion, give: the source fact, the rule you applied, the recommendation, and what remains uncertain. If a required fact is missing, ask before making the recommendation.

Test three different situations

Try an ordinary task, an incomplete task, and a conflicting task. Check whether the output format holds up and whether the AI stops where you intended. A method that only works on a carefully prepared demonstration is not ready for strangers to use.

Keep a short record of the input, what you expected, what happened, and the correction you made. Submit an update when the instructions change. Review should focus on the final candidate, because examples and dependency changes can introduce new behavior even when the overall job stays the same.