Quick Answer
- Most "magic phrase" prompting advice from 2023 has lost its edge. What still holds up is structural: show examples, request a specific format, and ask the model to check its own work
- Few-shot prompting (adding example input-output pairs) is the best-documented technique. In published studies it moved a clinical extraction task from an F1 score of 0.288 to 0.676 and a histopathology task from 40% to 90% accuracy
- Role prompting ("you are an expert...") has mixed results: some studies find no accuracy gain, and its dependable effect is on tone
Why Most Prompting Advice Doesn't Hold Up Anymore
Some techniques deliver measurable gains. Others, despite being widely repeated, do close to nothing for accuracy. A 2025 review of prompt research ("The Prompt Report") catalogued 58 text-based techniques, and a 2025 follow-up guide distilled it to six core ones: zero-shot, few-shot, thought generation, ensembling, self-criticism and decomposition.
The ten tips below come from the techniques with the clearest evidence, plus a few plain habits that cost nothing to try.
What Does the Research Actually Show?
- Few-shot prompting improved a clinical text extraction task from an F1 score of 0.288 to 0.676, more than double, in a published study
- A histopathology classification task jumped from 40% to 90% accuracy using 20-shot prompting in another published evaluation
- Self-consistency combined with chain-of-thought prompting gained 17.9 points on GSM8K, 11.0 on SVAMP and 12.2 on AQuA over chain-of-thought alone, according to a systematic survey of prompt engineering
- Role prompting ("you are a senior engineer") shows mixed results. One study found no significant change in accuracy on 2,410 MMLU questions, while another reported a gain on AQuA reasoning. Treat it as a tone and audience tool, not an accuracy fix
Which 10 Techniques Are Worth Using?
Backed by strong evidence
-
Give one or two examples of what "good" looks like. Few-shot prompting is the most consistently effective technique in the research above, and it's the easiest to actually do
-
Ask for the answer, then a self-critique of that answer. Self-criticism and self-reflection techniques improve output quality across studies
-
For anything involving reasoning or math, ask for the work to be shown, then checked twice. Chain-of-thought combined with self-consistency delivers measured gains on reasoning-heavy tasks specifically
-
Specify the exact output format you want. A table, a bulleted list or a word limit removes guesswork about what a good answer looks like
-
Match your examples closely to the actual task. An example that resembles your real input guides the model better than a loosely related one
Genuinely useful, less dramatically proven
-
"Ask me clarifying questions before answering." Targets the most common cause of a bad answer, a vague question
-
"Tell me when you're not sure, don't just sound confident." Doesn't eliminate errors, but surfaces the ones worth double-checking
-
"Give me the bottom line first, then the details." Respects your time and lets you judge whether you need the full explanation
-
Ask for the answer ranked by confidence. Turns a flat list of options into something you can prioritize
Weaker than its reputation suggests
- Giving it a role or persona ("explain this as a doctor would") still has a place, but it shapes tone and audience fit more than correctness. Don't rely on it to fix an answer that's actually wrong
How Do You Chain These Together?
Combined, these work better than alone. A realistic example:
"Here are two examples of the format I want. Ask me clarifying questions before answering. Once you have what you need, give me the bottom line first, then the details as a table. Flag anything you're not fully certain about."
Paste that into ChatGPT or Claude, then add your task. For how the two differ in prompting style, see how to prompt ChatGPT, Claude and Gemini differently.
Where Won't These Techniques Save You?
- They don't fix a genuinely unclear question. Structure helps the model organize a good answer, it can't manufacture information you never provided
- They don't replace fact-checking. Asking for confidence levels reduces false certainty, it doesn't guarantee the answer is correct. See why AI hallucinates for the failure modes
- Some techniques matter less on reasoning models. Explicit step-by-step requests add less on models that already reason internally by default; the structural techniques above still apply
For the bigger picture on whether prompt wording still matters, read Is Prompt Engineering Dead? and the system vs user prompts guide. Next step: take one prompt you use weekly and add an example, a format and a request to flag uncertainty.
Results vary by model, task and domain. Treat these figures as directional evidence, not guarantees for your use case.
