Directional Stimulus Prompting

Directional stimulus prompting (DSP) steers a large, often black-box model by injecting small, targeted hints — the "stimulus" — into the prompt, such as keywords the output should include or emphasize. Introduced by Li et al. (2023), its distinctive twist is that a much smaller, tunable policy model is trained to generate those hints automatically for each input, so you influence a frozen LLM's behavior without fine-tuning it. In the original work, generating keyword cues nudged summaries to cover the points that mattered, improving alignment with reference summaries. For SaaS builders, the everyday takeaway is the pattern itself: adding a compact directional cue ("emphasize pricing and integrations"; "include these entities") is a cheap way to bias output toward what you want without rewriting the whole instruction. The full trained-policy version is heavier engineering, worth it mainly when you're calling a closed API you can't fine-tune but still need controllable, on-brand output at scale. For simpler needs, hand-written stimuli capture most of the benefit.