Use AI to propose options, then select with explicit constraints and a small real-world test. Keep idea generation separate from evidence about whether an idea will work.
Example: choose a workshop activity
A library organizer needs a ten-minute opening activity for six adults. Participants have paper and pens. The activity must use no personal accounts and fit a €0 materials budget. The goal is to help people practice writing a clear instruction.
Propose three different ten-minute opening activities for six adults.
Goal: practice writing a clear instruction.
Available: paper and pens. Budget: €0. No personal accounts.
For each idea give steps, required materials and a way to check completion.
Then compare the ideas against these constraints.
Treat predicted participant reactions as hypotheses.Example idea set
| Idea | Activity | Completion check |
|---|---|---|
| Draw from instructions | Pairs describe a simple arrangement of shapes for a partner to draw | Compare the drawing with the intended arrangement |
| Repair a vague brief | Each pair rewrites “help with our event” into one concrete task | Another pair can identify the required deliverable |
| Sort facts and guesses | Pairs mark confirmed details and missing information in a short notice | Every asserted logistical fact has an input source |
All three fit the listed materials and avoid account setup. Choose “repair a vague brief” if the next lesson is about goals: it produces the artifact participants will use immediately. That is an instructional choice based on this brief, not a claim about measured enjoyment.
Reject an option for a concrete reason
A live competition requiring a new app account fails the account constraint. A thirty-minute role-play fails the time limit. Ask for a repair or remove the option rather than overlooking the constraint because the idea sounds creative.
OpenAI’s evaluation guidance recommends judging behavior with task-specific criteria. Here the checks are time, materials, account access and a visible learner artifact. You can use pass/fail decisions without inventing precise effectiveness scores.
Run a small test
Rehearse the instructions with a colleague and a timer. Check whether the setup, activity and review fit ten minutes. Record confusing instructions and revise them. Only then use actual participant feedback to judge the activity’s usefulness.
For practice, change one constraint and see which idea remains viable. For a commercial proposal, use the business-validation lesson to distinguish customer hypotheses from market evidence.
The Poe overview explains a platform for accessing multiple bots. If comparing candidates from different models, use the same constraints and record the actual differences you observe.