What Prompt Research Supports

Prompt advice is rich in folklore. A smaller set of findings has enough empirical support to guide system design. The figures below are the claims reported by the cited research.

Match instruction style to the task

Declarative descriptions of the desired end state outperform imperative step lists on some complex reasoning tasks, with SatLM reporting roughly a 23 percent gap (arXiv:2305.09656). Imperative instructions remain useful for sequential tool execution.

Describe outcomes for reasoning work. Give ordered steps for mechanical work.

Add detail selectively

The DETAIL study (arXiv:2512.02246) found that added detail improved math and code tasks, had little effect on decision-making, and harmed creative work. Specify boundaries, output format, and success criteria. Leave room in the approach when exploration is part of the task.

Prefer positive behaviour

Large lists of negative constraints repeatedly activate the concepts they prohibit. Where possible, state the behaviour you want. “Use dependency injection” gives the model a clearer path than “never use global state.”

Security boundaries are different. Consequential prohibitions still belong in the system, but enforcement should live in permissions and controls rather than prompt wording alone.

Keep personas minimal

A study covering 162 personas and 2,410 questions found no general factual or reasoning benefit from role-play backstories, with accuracy declining in roughly 30 percent of tested cases (arXiv:2311.10054). A short focus statement such as “focus on security implications” usually provides the useful part without the narrative overhead.

Use reasoning traces deliberately

Chain-of-thought prompting changes how a model approaches a task, but it does not guarantee valid reasoning. On modern reasoning models, forcing visible step-by-step output can add latency without improving accuracy. Use explicit reasoning traces for debugging, review, or required transparency, not as a universal default.

Put important material where it will be noticed

Order matters. Put the task and critical constraints early. In long prompts, repeat only the load-bearing success conditions near the end. Avoid burying important requirements in the middle of retrieved material.

These are starting hypotheses, not permanent laws. Test them against the models and tasks you actually run. The next chapter explains how to manage those tests and prompt changes over time.