Prompt engineering: a practical guide
The term is arguably already outdated — here's what actually works today, and why ChatGPT, Claude, and Gemini don't all want to be prompted the same way.
Why "prompt engineering" is already a shrinking title
In June 2025, AI researcher Andrej Karpathy argued in a widely-shared post that "prompt engineering" undersells what the job actually is. His framing: the model is like a processor, its context window is like working memory, and the real skill is deciding exactly what information belongs in that limited space — instructions, examples, retrieved data, prior conversation — so the model has what it needs and nothing that distracts it. That reframing lines up with what's happened to the job market for it: postings using the exact title "prompt engineer" dropped roughly 40% between 2024 and 2025, not because the skill stopped mattering, but because it folded into broader AI workflow and automation design roles instead of standing alone.
None of that makes the underlying skill less useful for an everyday user — if anything, it's more useful, because it means treating a prompt less like a magic phrase and more like a clear, well-scoped request. The techniques below are the ones that consistently show up across current guidance from AI labs and practitioners, kept to the ones actually worth remembering rather than a long list of jargon.
The techniques actually worth knowing
Zero-shot
Just ask clearly, with no examples. Works well for straightforward tasks on capable models — try this first before adding complexity.
Few-shot
Show one example, and only add more if the output still misses the mark. Most guidance now says start with one, not several.
Chain-of-thought
Ask the model to reason step by step before answering — helps noticeably on math, logic, and multi-part questions.
Role prompting
Give a brief, task-relevant role at the start. Keep it short — an elaborate persona tends to add noise, not accuracy.
Prompt chaining
Break a big task into sequential prompts, feeding each result into the next step, instead of one giant request.
Self-consistency
For tricky reasoning tasks, generate a few independent answers and go with the one they mostly agree on.
What a stronger prompt actually looks like
Weak prompt
- "Write about our product launch."
- No audience, tone, length, or goal specified — the model has to guess all four.
Stronger prompt
- "Write a 150-word announcement for our email subscribers about [product], in an upbeat but not salesy tone, ending with a single clear call to action."
- Audience, length, tone, and goal are all explicit — nothing left to guesswork.
One more habit worth adding to almost any factual prompt: explicitly give the model permission to say it isn't sure, instead of guessing. Something as simple as "if the data doesn't support a clear answer, say so" measurably reduces confidently wrong answers.
The same prompt, different models
A prompting style that works well on one AI Chat tool won't necessarily get the same result on another — each has its own tendencies, based on current guidance from the companies building them.
Mistakes that quietly waste good prompts
Vague success criteria
- If you can't describe what a good answer looks like, the model can't aim for it either.
One giant request
- Multi-stage tasks crammed into a single prompt tend to lose coherence — break them into steps instead.
Never testing again
- A prompt that worked once can quietly stop working after a model update — re-check prompts you rely on regularly.
A short checklist to start with
Before sending a prompt for anything that matters, it's worth running through a short list: is the goal explicit, not implied? Is there a clear audience, tone, and length where relevant? Would one example help more than a paragraph of instructions? And for anything factual, have you given the model permission to say "I'm not sure" instead of guessing? Getting comfortable with these habits matters more than memorizing technique names — the AI Glossary covers the terminology if you want the vocabulary too, and the tools in AI Chat are the fastest place to practice all of this directly.