What Actually Makes a Prompt Work

1. Specificity beats everything

Every vendor's documentation converges here, and it's the one recommendation no study has contradicted. Meta's own guidance puts it plainly: be clear and concise — your prompt should be easy to understand and provide enough information for the model to generate relevant output, avoiding jargon or technical terms that may confuse the model.

The practical version: "Write a marketing email" gives the model nothing to work with. "Write a 150-word marketing email promoting a project-management app to freelance designers, casual and encouraging in tone, ending with one clear call-to-action" gives it five constraints it can actually satisfy.

This is not a trick. It's just the difference between a brief and a wish.

2. Structure helps — but not for the reason people think

Separating your prompt into labelled sections — the task, the context, the constraints, the audience — makes a real difference. But the reason is more interesting than "AI likes XML."

A 2026 study ran 9,649 experiments across 11 models and 4 formats (YAML, Markdown, JSON, and a compact format called TOON), with schemas ranging from 10 to 10,000 items. The finding: format does not significantly affect aggregate accuracy (chi-squared = 2.45, p = 0.484), though individual models — particularly open-source ones — show format-specific sensitivities.

So the brackets themselves aren't magic. What structure actually does is force you to fill in the sections you'd otherwise leave blank. You can't skip the audience field if there's an audience field staring at you. Structure is a completeness checklist wearing a syntax costume — and completeness is what moves the needle.

There's a second, real benefit: labelled sections make it unambiguous which part of your message is instruction and which part is data you're handing over. That distinction matters a great deal when you're pasting in someone else's text.

3. Examples are the strongest single lever

If you can show the model one or two examples of what you want, do it. This is called few-shot prompting, and it's the most consistently effective technique across every model family.

Google's documentation demonstrates the mechanism directly — a prompt providing two examples that showed preference for shorter explanations guided the model to choose the shorter explanation, as opposed to the longer one it had picked previously, and notes that models can often pick up on patterns from a few examples, though you may need to experiment with the number.

You don't need a perfect example. An imperfect one you edited into shape works fine — and is usually faster to produce than three paragraphs describing the shape you want.

4. Say what format you want, explicitly

This is the one that surprised researchers. In a Wharton study testing which prompt modifications actually change outcomes, prompt modifications like politeness influenced individual responses but had minimal overall effect, and aggregate model characteristics dominated over specific prompting strategies — while removing formatting instructions led to performance drops.

Read that asymmetry carefully. Politeness didn't matter. Removing the format instruction did. If you're going to spend one sentence on prompt engineering, spend it on the output format, not on being nice.

Google's guidance is the same: you can give instructions specifying the format of the response — a table, a bulleted list, an elevator pitch, keywords, a sentence, a paragraph.

5. Put your instructions in the right place — which depends on the model

Here's a rule almost nobody knows, straight from Google's Gemini documentation: when working with large datasets such as entire books, codebases, or long videos, place your specific instructions or questions at the end of the prompt, after the data context, and anchor the model's reasoning by starting your question with a phrase like "Based on the preceding information…".

Meanwhile, most Claude and GPT guidance puts the task first. Both are correct — for their model. This is the clearest example of why "one universal prompt" is a myth, and why PromptMe generates model-specific variants rather than one output.

❌ "Act as an expert" doesn't make the AI smarter

The single most-repeated piece of prompting advice is to assign the model an expert persona. It has been tested, repeatedly, and it does not do what people think.

A peer-reviewed study presented at EMNLP 2024 evaluated 162 roles covering 6 types of interpersonal relationships and 8 domains of expertise, across 4 popular LLM families and 2,410 factual questions, and found that adding personas in system prompts does not improve model performance. It also found that while aggregating results from the best persona for each question significantly improves accuracy, automatically identifying that best persona is challenging, with predictions often performing no better than random selection.

A 2025 Wharton replication on much harder, PhD-level questions reached the same conclusion. As Ethan Mollick, one of its authors, put it: "We found that telling the AI 'you are a great physicist' doesn't make it significantly more accurate at answering physics questions, nor does 'you are a lawyer' make it worse" — adding that roles can be helpful for framing context and style, but they aren't magical, and sometimes giving the AI a role can actually lower accuracy.

So should you ever use a persona? Yes — for voice. "Write this the way a friendly veterinarian would explain it to a nervous pet owner" is a genuinely useful instruction, because you're specifying tone and vocabulary. Just don't expect it to improve factual accuracy. Use personas to shape how it says something, not to make it know more.

Notably, this advice still appears in vendor documentation. The Wharton team pointed this out directly: Google's Vertex AI prompt-design guide says to "Assign a role", Anthropic's samples include "You are an expert AI tax analyst", and OpenAI's developer materials take a similar approach — while independent research on performance gains from persona-based prompting is rare and inconsistent.

❌ "Think step by step" is mostly obsolete on modern models

Chain-of-thought prompting — telling the model to reason through a problem before answering — was a genuine breakthrough in 2022. In 2026, most frontier models do it automatically.

Wharton tested it directly and found that its effectiveness varies significantly by model type and task: non-reasoning models show modest average improvements but increased variability in answers, while reasoning models gain only marginal benefits despite substantial time costs of 20–80%.

Their bottom line: a simple chain-of-thought prompt is generally still useful for boosting average performance in non-reasoning models, especially older or smaller ones, but for dedicated reasoning models the added benefit appears negligible and may not justify the substantial increase in processing time.

Practical rule: if you're using a reasoning model (Claude with extended thinking, GPT-5.x at medium or high reasoning effort, Gemini 3 Thinking, DeepSeek with thinking enabled), skip it — you're paying for tokens and latency to get something the model already does. If you're using a fast, cheap, non-reasoning model, it can still help.

❌ Politeness doesn't reliably help (or hurt)

Please and thank you are fine. They're just not a technique. The Wharton contingency study found being polite or commanding yields question-specific differences rather than global improvements, and these effects often diminish when aggregated. Be polite because you want to be, not because you think it's buying accuracy.

❌ There is no universal "best prompt"

The most-cited finding from the same study is the one nobody wants to hear: prompt variations produce inconsistent effects, challenging the notion of universally effective prompting techniques, and emphasising the necessity of context-specific evaluation.

The same prompt can gain accuracy on one question and lose it on the next. Which means the honest advice isn't "use this template" — it's "try it, look at what came back, and adjust." Iteration beats optimisation.

xAI's own developer guidance makes this point about their fast models: instead of spending twenty minutes crafting the "perfect" prompt, fire off a quick attempt and refine based on the results — this iterative approach often gets you to a better outcome faster than trying to nail it in one shot.