The EmotionPrompt study reports positive results. For your workflow, separate encouragement from a task-instruction change and score saved outputs against the same criteria.
In this guide
What did the EmotionPrompt paper find?
Li and colleagues studied emotional additions to prompts across 45 tasks and several models, including ChatGPT and GPT-4. Their 2023 paper reports relative improvements of 8.00% on Instruction Induction and 115% on BIG-Bench in its tested settings.
Those are relative benchmark changes, not percentage-point gains or a promised improvement in today's product descriptions. The paper uses different task metrics; its BIG-Bench result is not an accuracy score of 115%. The result shows that wording can matter, without establishing that the model feels encouragement.
How should you judge the output rather than the encouragement?
These three sentences are authored grading examples, not model outputs or experiment results. They show the acceptance rule to apply to each real run. When testing encouragement, keep the task instructions unchanged and add only the emotional wording. Compare a clearer-instruction variant separately, or use four variants to distinguish those two changes. Record repeated runs and model settings before drawing a conclusion.
| Authored output for the same source note | Condition retained? | Reviewer decision |
|---|---|---|
| Maya will deliver the draft Friday. | No: access must return Wednesday | Revise the commitment |
| Maya will deliver an excellent draft Friday. | No: positive wording changed no factual condition | Revise the commitment |
| Maya can draft by Friday if access returns Wednesday; confirm calendar dates. | Yes | Meets this practice criterion |
How can you separate encouragement from clearer instructions?
- Save representative real inputs you are allowed to use, including a case with a condition the answer must retain.
- Compare the original instructions with the same instructions plus one emotional sentence. Keep the model, tools, input, and available generation settings the same.
- Test clearer task instructions separately. For four variants use original, original plus encouragement, clearer instructions, and clearer instructions plus encouragement.
- Save repeated outputs without choosing only the best response. Review them in mixed order against the same factual and task-completion criteria.
- Record model/date/settings, corrections needed, and the reason for each decision. Decide whether the observed difference is worth retaining for this job.
Keep a comparison record you can rerun
Use one row per output. The three practice sentences above demonstrate grading; they are not experiment results to enter into your trial. Keep cost and latency beside quality when they affect whether the workflow is worth repeating.
case_id: meeting-access-01 source: T02, draft conditional on access returning Wednesday model_and_date: record actual values variant: original / encouragement / clearer / clearer-plus-encouragement run_number: record every repeated run output_file: save unchanged response criterion: access condition retained; calendar dates requested result_and_reason: complete after inspecting the output
Which change should you keep?
Keep a change when it repeatedly helps the work you measured, and retain examples that still fail. A positive paper result is a reason to test emotional wording. It does not replace the source evidence, required output, or feedback about a specific error.
Recheck after the model, input type, or intended audience changes. If encouragement only changes friendliness while the same condition is still missing, improve the instructions and evidence pack before spending more effort on the motivational phrase. Record that as a result for this task rather than a universal claim about all AI.