Here is a party trick that turned out to be a real research finding. Take any prompt you would send to a language model and staple one extra sentence onto the end: "This is very important to my career." That sentence tells the model nothing about the task. It does not hint at the answer, add a rule, or narrow the question. And yet, on open-ended work, the reply often comes back longer, more careful, and better organised than the plain version.
That is Emotion-Prompting, introduced by Li et al. in 2023 ("Large Language Models Understand and Can Be Enhanced by Emotional Stimuli"). They appended feeling-loaded phrases to otherwise-normal prompts across dozens of tasks and several models, and measured a consistent bump on average, with double-digit relative gains on some generative tasks. Same model, same task, same temperature. The only difference was a sentence of manufactured stakes.
Why on earth would that work
The model does not have feelings, so you are not motivating anything. The leading explanation is boring and statistical: in the training data, stakes and urgency co-occur with care. The emails people flagged as important were the ones they wrote carefully. The passages that said "this matters" were followed by thorough, well-structured continuations. The model learned that correlation. When you append "this is important to my career," you shift the conditional distribution toward the region where high-effort human writing lives. You are not pep-talking a mind. You are steering the sampler into a better neighbourhood of the text it already knows how to produce.
The three flavours
The paper's stimuli cluster into three families, and it is worth knowing which one you are reaching for:
- Social-esteem: invoke other people's judgment or your standing. "This is very important to my career." "Others are counting on this."
- Urgency and stakes: raise the cost of getting it wrong. "Getting this right really matters."
- Self-monitoring: tell the model to check and pace itself. "Are you sure that is your final answer?" "Take a deep breath and work through it step by step."
That last family tends to be the strongest, and it is not hard to see why: "step by step" is doing double duty as a reasoning cue. It overlaps with chain-of-thought, not just emotion.
Where it helps and where it is useless
Emotion-Prompting pays off when "more effort" has somewhere to go: writing, summarising, brainstorming, advice, explanation, subjective reasoning. Anywhere a longer, more careful answer is genuinely a better answer, the stimulus has headroom to work with.
Where the answer is fixed and singular, it does nothing. 17 times 24 is 408 whether or not your career depends on it. A lookup, a format conversion, a strict yes/no classification against a definite key — there is no "harder" to try, so the stimulus just adds words around the same result, and can even inject distracting hedging. Save it for quality tasks; reach for a calculator or a tool on deterministic ones.
The part people skip
Not every emotional sentence points the same direction. Encouragement and self-monitoring tend to help. Threats do not. "Get this wrong and you will be fired" is a coin-flip that can push the model toward terse, hedged, blame-avoiding output — worse than the plain answer. When in doubt, encourage; do not threaten.
And the whole effect is fragile. The biggest gains showed up on earlier, smaller, less heavily-aligned models. Strong, modern, instruction-tuned models already write high-effort answers by default, so there is less slack to recover, and the measured lift often shrinks toward noise. It is also brittle to phrasing — swap a word and it moves. A boost you measured last quarter on last quarter's model may simply be gone today.
How to actually use it
Treat the stimulus as a single tunable knob, not an article of faith. Pick one phrase that fits the task. Run it against a held-out eval set with and without. Keep it only if it moves your real metric by more than noise, and pin the winner to a regression test so a model update cannot quietly erase it. Do not stack five stimuli hoping they add up — returns diminish fast and a bloated emotional preamble starts to distract. One well-chosen line beats a pile of them.
It is the smallest technique in the whole prompt-engineering toolbox: one sentence, no loop, no tool, no second model. Cheap to try, easy to measure, and never guaranteed. Just remember you are steering a distribution, not talking to someone who cares.
I built an interactive version — toggle stimuli, watch a plain answer and an emotionally-charged one diverge, and see one phrase that does nothing and one that backfires: https://dev48v.infy.uk/prompt/day31-emotion-prompting.html
Top comments (0)