OpenAI released the Decisions API beta on 6 October 2026. Its current guide lists gpt-6-luna and typed questions for conditions, choices and scores. This developer API is not a general writing chat. Before choosing a tool or composing a prompt, decide what kind of answer the task actually needs.
Choosing which supplied book description fits a stated theme can be a fixed-choice task. Writing four discussion questions and a timetable is a generative task. The example below is for a writing assistant; do not expect the Decisions endpoint to produce that whole plan.
A helpful prompt tells the assistant what success looks like. “Make a plan” leaves it to invent the goal, time available and format. You do not need a long formula; you need the few facts that change the answer.
Give the missing decision points
For a fictional book club, try: “Use these three book descriptions to prepare a 30-minute discussion for six adults. Give four questions and a simple timetable. Do not add facts about books that are not in the descriptions.”
The task, inputs, audience, duration and boundary are visible. The numbers are our example, not a promise about an assistant's performance.
State how to handle uncertainty
If a description omits the ending, ask the assistant to mark that gap rather than assume one. Keep instructions distinct from source text. A quoted document may contain wording to analyse, not commands to follow.
Refine one failure at a time
Inspect the first result. If the timetable exceeds 30 minutes, ask for that correction. If a question relies on an invented event, point to the relevant description and remove the unsupported premise. Keep a working prompt once it succeeds. Repeated tasks may benefit from a skill later, but a specific request and an actual result check are the foundation.