Turn a spoken brainstorm into an organized idea list
A spoken brainstorm easily mixes possibilities with decisions. Separate generating ideas from evaluating them, then make an explicit selection. The goal is a usable candidate list, not the largest number of things an assistant can say.
Start with a boundary that creates useful ideas
Name the problem, audience and constraint before asking for suggestions. “Make onboarding better” can produce anything from a new video to a rebuilt product. “Help first-time workshop attendees find the right room using material we can change this week” defines a problem you can actually explore.
Keep observations separate from explanations. “Three attendees asked the receptionist for directions” is an observation in this invented example. “People dislike our signage” is a hypothesis. If you present the hypothesis as a fact, the assistant may optimize for a problem you have not established.
Round one: generate contrasting approaches
Use this prompt: “Suggest four substantially different ways to help a visitor reach Room C. We can edit the confirmation email and print signs, but cannot change the building. Label each idea with the assumption it depends on. Do not select a winner yet.” Speak your real constraints slowly enough to review their transcription.
Illustrative candidates might include an annotated arrival photo, a short landmark-based instruction, a color-coded route and a staffed welcome point. The last candidate conflicts with a no-staff constraint if you have one; identifying that mismatch is part of the exercise, not proof that the model has checked feasibility for you.
Round two: compare on the same criteria
Choose criteria before seeing a favorite answer. For the room example, compare preparation time, accessibility of the instructions and whether the visitor can use them without mobile data. Mark unknowns rather than inventing measurements. A model-generated “90% improvement” has no place in the comparison unless you supplied a valid measurement.
- Read the candidate names back and remove duplicate approaches.
- Reject ideas that violate a hard constraint.
- Give each remaining idea the same three criteria.
- Ask what evidence would change the ranking.
- Select a small trial and name who will check it.
Round three: turn the choice into a test
For the annotated-photo candidate, an original trial could be: print the arrival photo, give it to a colleague unfamiliar with the route, and ask them to describe the first turn without coaching. Record the confusing landmark and revise the image. This produces concrete feedback rather than an imaginary conversion rate.
The assistant can help draft the trial instructions when you ask it. You remain responsible for running the trial and recording the result. GPT Realtime Assistant provides a voice session and a retained conversation; this workflow does not imply a built-in brainstorming board, voting system or automatic experiment runner.
Keep a short decision record
Finish by manually noting the selected idea, rejected alternative, reason and next check. Ask for a recap of the discussion if helpful, but compare it against what you actually chose. Keep speculative ideas labeled as speculative so a later reader does not confuse an energetic conversation with approved work.
If colleagues join the brainstorm, agree before capturing their speech. Allow a person to contribute without being recorded, and keep confidential proposals out of cloud processing unless your organization permits it. A productive voice exercise ends with one thing to try and a clear way to decide whether it helped.
