How does prompt design alter what kind of creativity LLMs can express?

This explores whether prompt design changes the *kind* of creativity an LLM can express, not just how well it performs — and the corpus suggests the lever is real but blunt, because most prompting nudges the model toward the safe, feasible end of the creative spectrum rather than the genuinely novel one. The starting point is that creativity isn't one thing: one line of work argues creative reasoning splits into three distinct modes — combinational (recombining known ideas), exploratory (searching within a space), and transformational (breaking the space open) — and that current LLM methods only ever exercise conventional problem-solving, leaving the transformational mode essentially untouched Can LLMs reason creatively beyond conventional problem-solving?. So before asking how prompts steer creativity, it helps to know there are several creativities to steer toward.

Where it gets interesting is that the most common prompting moves seem to trade novelty away. LLM-generated design concepts score *higher* on feasibility and usefulness but *lower* on novelty than human crowdsourced ones — and few-shot prompting, the standard "here are some examples" technique, makes this worse: it improves quality alignment while actively shrinking diversity Why do LLMs excel at feasible design but struggle with novelty?. Yet the raw capacity for novelty is clearly there: in a large study of NLP researchers, LLM-generated research ideas were rated *more* novel than expert human ideas, precisely because expert knowledge constrains exploration while the model roams wider conceptual territory Do language models generate more novel research ideas than experts?. Put those two together and a picture emerges — the model can be wide-roaming or safe, and prompt design (especially exemplars) is one of the dials that decides which.

The surprising part is how *little* of this is about meaning and how much is about statistics. Semantically identical prompts produce systematically different outputs because the model responds to how frequently a phrasing appeared in pre-training, not to what it means — higher-frequency wordings win Why do semantically identical prompts produce different LLM outputs?. That implies a high-frequency, conventional phrasing may quietly pull the model toward conventional output, while an unusual framing opens a different region of its distribution. Tone does something similar: emotional framing shifts which information surfaces Does emotional tone in prompts change what information LLMs provide?, and appending motivational phrases like "this is very important to my career" reliably changes performance through framing alone, not new information Can emotional phrases in prompts improve language model performance?. The prompt isn't just a question — it's a coordinate that lands the model somewhere in its space.

Two more findings complicate the easy story that "better prompt = more creativity." First, technique doesn't transfer cleanly: step-by-step reasoning helps cheap models but *reduces* accuracy in strong ones, and the right move depends on task structure rather than generic best practice Do prompt techniques work the same across all LLM tiers?; relatedly, chain-of-thought only helps when the question's content flows into the prompt before reasoning starts — for simple questions it backfires Why do some questions perform better without step-by-step reasoning?. Forcing structured reasoning where it doesn't fit can suppress the looser association that creative output needs. Second, the model resists being reshaped: most open LLMs cling to a default "ENFJ-like" personality and refuse prompted personas, so prompt-driven creative range has a ceiling set by what the model already is Can open language models adopt different personalities through prompting?.

If you want to go deeper on the design side rather than the model side, two notes reframe prompting as a structured craft: prompt quality can be decomposed into six measurable dimensions (communication, cognition, instruction, logic, hallucination, responsibility) where improving one cascades into others Can we measure prompt quality independent of model outputs?, and there's a sharp warning that iterative, ad-hoc prompt tweaking introduces hidden bias and self-fulfilling loops — meaning the very process of "prompting until it's creative" can manufacture the creativity you were hoping to measure Does iterative prompt engineering undermine scientific validity?. The takeaway the reader may not have expected: prompt design alters creativity less by *adding* imagination and more by selecting which region of an already-fixed distribution the model speaks from — and the default pull of most techniques is toward the feasible, not the novel.