A beginner's guide to prompt engineering that actually holds up

Ninety percent of prompt advice on the internet is a list of magic phrases. "Take a deep breath." "You are a world-class expert." "I will tip you $200." Some of these did something on some model in some month of some year. None of them are a skill.

The actual skill is smaller and more boring than the content industry wants it to be. Here it is.

Move one: give the model a role

Models are prediction engines that behave differently depending on the register you put them in. "Explain recursion" gets you a Wikipedia paragraph. "You are a patient tutor explaining recursion to a bright 14-year-old who has just learned about loops" gets you something usable.

The role sets vocabulary, depth, and default assumptions in one line. It is the cheapest quality upgrade available.

Move two: state the task in one unambiguous sentence

Not "help me with my CV." That is a topic, not a task. "Rewrite the three bullet points below so each one leads with a measurable result" is a task. If you cannot say what a correct answer looks like, the model cannot produce one.

Move three: add constraints, especially negative ones

Constraints are where output goes from plausible to right:

• Length — "under 120 words", not "be concise". Concise is subjective; 120 is not.
• Audience — who reads this changes everything.
• Tone — pick two adjectives. "Direct and warm." "Formal and neutral."
• Never do — the underused one. "Never invent statistics." "Never use the word 'delve'." "Never apologise."

Move four: specify the output shape

Say exactly what comes back. "Return a markdown table with columns Name, Cost, Trade-off." "Return only the rewritten text, no explanation." Models default to conversational padding — an intro, a summary, an offer to help further. If you do not want that, ban it explicitly.

Then: iterate like an engineer, not a gambler

Here is the part that separates people who get good from people who keep collecting templates.

Change one thing at a time. If you rewrite the whole prompt and the output improves, you have learned nothing about why.

Keep the failures. When output is wrong, the failure tells you which instruction was missing. Too long? You never set a length. Made something up? You never banned invention. Bad output is diagnostic, not random.

Give an example. One good input/output pair in the prompt (one-shot prompting) beats three paragraphs of description. Two examples beat one. This is the highest-leverage move most beginners never make.

Reuse what works. When a prompt earns its keep, save it. A personal library of eight prompts you trust is worth more than a thousand-item swipe file. Our prompt generator is built on the same idea — structured recipes for building, creative work, and dev productivity, ready to edit.

What about long context and reasoning models?

Two adjustments. With long context, put the instructions after the pasted material as well as before — attention frays across very long inputs, and a closing restatement of the task measurably helps. With reasoning models, stop telling them how to think; "think step by step" is now largely redundant, and over-scripting the process can make results worse. Describe the destination, not the route.

The uncomfortable truth

There is no phrase that unlocks a secret better model. There is only clarity: who, what, within which limits, in what shape. Get precise about those four and your results improve on every platform, this year and next.

Unclear prompt in, unclear answer out. Get specific, and get it ticked.