Key answer
Templates mainly help reuse task structure and remind you to supply information. Evidence does not establish fixed wording that stays optimal for every model and task. Ease of expression and higher model scores require separate judgments.
Full explanation
Sclar et al. compared prompts with similar meanings but different formats across over 50 tasks and several models, observing substantial differences and only weak correlations between models' preferences. Formatting can matter without transferring across models; one high-scoring prompt does not reveal a universal template. The initial work mainly involved then-current Llama models and GPT-3.5, so its largest gaps do not predict arbitrary tasks today. Original study
Google presents templates as starting points and emphasizes task-specific iteration. This helps organize work without establishing a universal percentage gain. Official guide
For beginners, a removable skeleton is more useful than a long form that must always be filled:
Complete [task] for [purpose/audience], using [necessary material], observing [key constraints], and deliver [required result].
Omit already-clear parts. Do not invent background when none exists or add roles, examples and self-checks to everything. Retain empirically useful versions for recurring work; compare and update when models, tasks or observed errors change.
Research cannot guarantee future usefulness. Particular wording and format advantages may change. As long as tasks include private goals and tradeoffs AI cannot know, some communication remains necessary. That information requirement does not require today's templates to remain unchanged.
References
2. Google, Prompt design strategies, official Gemini developer documentation