There’s a strange industry built on prompting advice. Threads promising “10 secret prompts,” courses selling magic phrases, templates traded like recipes. Almost none of it cites evidence, and some of the most repeated tricks measurably stopped working a model generation ago. Meanwhile, the techniques that do have research behind them fit on one page and mostly amount to something unglamorous: being specific, showing examples, and giving the model room to reason.
This article covers only techniques with published research or official vendor documentation behind them, notes where the evidence has shifted, and skips the folklore entirely.
Which Prompting Techniques Are Actually Proven?
The techniques with the strongest evidence are: giving examples of what you want (few-shot prompting), asking the model to reason step by step before answering (chain-of-thought), writing specific instructions with relevant context, and breaking complex tasks into stages. Each has peer-reviewed research behind it, and each also has documented limits that the advice industry usually leaves out.
Chain-of-Thought: Strong Evidence, With a Twist
The most-cited prompting technique of all is chain-of-thought: prompting the model to work through intermediate reasoning steps before giving a final answer. The original 2022 research from Wei and colleagues showed dramatic gains on math and reasoning tasks, and a follow-up finding showed even the bare phrase “let’s think step by step” improved results without any examples. This is real, replicated, and foundational.
The twist is what’s happened since. Newer models with built-in reasoning already think step by step internally, and recent research revisiting chain-of-thought finds that on such models, pasting in hand-written reasoning examples can actually hurt performance compared to just asking the question plainly. The technique didn’t stop being true; the models absorbed it. The practical rule today: on a reasoning-enabled model, ask clearly and skip the “think step by step” theater. On a standard model, the phrase still earns its keep on genuinely multi-step problems.
Few-Shot Examples: The Most Underused Real Technique
Showing the model two or three examples of exactly what you want, format, tone, level of detail, remains one of the most reliably effective moves available, especially for output that needs a consistent structure. It’s also the technique casual users skip most, because writing examples feels like more work than describing what you want. It usually isn’t: one good example replaces a paragraph of adjectives, and the model imitates structure far more faithfully than it follows abstract style instructions.
The documented caveat: example quality and consistency matter more than quantity. Mismatched or sloppy examples teach the mismatch.
Specificity and Context: Boring, Documented, Effective
Every major vendor’s official prompting documentation converges on the same unglamorous advice: state the task precisely, include the context the model can’t know, say who the output is for, and define what good looks like. The comprehensive academic survey of the field, The Prompt Report, catalogs dozens of named techniques, and the pattern across them is that most workable ones are structured ways of being clearer, not incantations.
A practical version of “specific” that consistently pays off:
- State the role of the output, not the model: “a summary for an executive who hasn’t read the report” beats “act as an expert summarizer.”
- Include the constraint that actually matters: length, format, what must be excluded, what the output will be used for.
- Paste the relevant context rather than assuming the model knows your project, your audience, or your standards.
- Define the failure mode you’re worried about: “if a figure isn’t in the document, say so rather than estimating” prevents the specific error you care about.
- Ask for the output format explicitly, and if it’s structured, show it once.
What Doesn’t Have Good Evidence
| Popular advice | Status |
|---|---|
| “Act as a world-class expert” role-play openers | Weak: minor or no measured effect on factual quality in most modern testing; formatting the task well matters far more |
| Offering the model tips, threats, or emotional stakes | Folklore: occasionally shows tiny effects in one-off tests, doesn’t replicate reliably, not worth building habits on |
| Magic phrasing and “secret prompts” | Folklore: no mechanism, no replication; what works transfers poorly between models anyway |
| Hand-written step-by-step examples on reasoning models | Overturned: can now reduce accuracy versus a plain, clear question on reasoning-enabled models |
| Repeating instructions multiple times for emphasis | Weak: clarity once beats repetition; excess repetition wastes context and can distort attention |
The Meta-Technique: Test on Your Own Tasks
The honest limit of all prompting research is that results vary by model, task, and version, and vendors update models constantly. The habit that outlasts every individual technique is keeping a small set of your real tasks and re-testing your standard prompts when you switch models or a model updates, the same discipline covered in our guide to trialing an AI tool properly. Prompting is closer to an empirical craft than a spellbook: evidence tells you where to start, your own tests tell you what’s true for your work today.
It’s also worth knowing when prompting is the wrong lever entirely. If outputs are inconsistent run-to-run, the issue may be a settings problem, not a wording problem, and if a task needs behavior no prompt can hold reliably at volume, that’s the boundary where fine-tuning starts to make sense.
Frequently Asked Questions
Does “let’s think step by step” still work? On standard models, yes, for genuinely multi-step problems. On newer reasoning-enabled models that deliberate internally by default, it adds little, and pasting hand-written reasoning examples can measurably hurt compared to asking plainly.
How many examples should a few-shot prompt include? Two or three consistent, high-quality examples usually capture most of the benefit. Consistency between the examples matters more than the count, since the model imitates whatever pattern they share, including their flaws.
Do role prompts like “you are an expert lawyer” improve accuracy? The evidence is weak. Roles can nudge tone and framing, but modern testing generally finds task clarity and context drive quality, not the persona. Use a role to set voice if you like; don’t rely on it for correctness.
Why does the same prompt work differently on different models? Models are trained differently and updated often, so technique effectiveness genuinely shifts between them, which is exactly what happened to chain-of-thought examples on reasoning models. This is why testing on your own tasks beats memorizing tips.
Is there one prompt structure that works for everything? No, and distrust anyone selling one. The durable structure is closer to a checklist than a template: task, context, audience, constraints, format, and the failure mode to avoid, filled in with the specifics of your actual job.
What to Do With This
Delete the folder of magic prompts. Keep four habits: show an example instead of describing the style, state the context and the constraint that matters, name the failure you want avoided, and on reasoning models, just ask the question clearly. Then test on your own recurring tasks, because that’s the only benchmark that was ever about your work. Our practical coverage of getting more from these tools lives under productivity.