Directional Stimulus Prompting
Guiding an AI toward a desired output by injecting small hints, keywords, or cues into the prompt.
Also known as: DSP, Hint-Based Prompting, Stimulus Prompting
Category: Techniques
Tags: ai, prompting, llm-techniques, optimization
Explanation
Directional stimulus prompting (DSP) nudges a language model toward a specific kind of response by adding targeted hints - keywords, constraints, or stylistic cues - alongside the main request. Rather than rewriting the whole instruction, you provide a small 'stimulus' that steers generation without overspecifying it.
The technique was introduced in research showing that a small tunable model can learn to produce hints for a larger frozen model, but the core idea applies manually as well: a well-chosen word or phrase in the prompt measurably shifts outputs.
Examples:
- **Summarization**: Instead of 'Summarize this article,' add a hint listing key entities or themes: 'Summarize this article (hint: focus on financial impact, regulatory risk, timeline).'
- **Style control**: 'Write a product description (hint: confident, minimalist, no superlatives).'
- **Disambiguation**: For ambiguous questions, add domain hints: 'Explain Python (hint: programming language, not snake).'
- **Structured output**: 'List steps (hint: numbered, action verbs, one line each).'
Why it works:
- Hints activate relevant parts of the model's training distribution without rewriting the task.
- Small stimuli preserve the model's flexibility while constraining the output space.
- Keywords can outweigh full sentences because attention concentrates on distinctive tokens.
DSP is especially valuable when:
- You need many variations that each follow a different angle.
- Full instruction rewrites would be verbose or brittle.
- You want to externalize control (hint generation) from core instruction design.
It overlaps with few-shot prompting and role prompting but is lighter weight - a nudge rather than a full example or persona.
Related Concepts
← Back to all concepts