Prompt Engineering with ChatGPT

What makes a prompt reliable

A reliable prompt removes ambiguity about four things: what to do, for whom, in what form, and under what rules. If any of those is missing, ChatGPT fills the gap with a guess, and different guesses on different days produce different results. Reliability does not come from length; it comes from answering those four questions plainly, every time.

Test reliability by running the same prompt on several different inputs. If the output format changes, the tone drifts, or details appear that were not in the input, the prompt needs tightening, not a better model. Most prompt problems are specification problems, and they are fixable in the prompt itself.

Structure: role, task, context, format, constraints

A practical structure has five parts. Role: who ChatGPT should act as. Task: what to do, in one or two sentences. Context: the facts, audience, and background it needs. Format: the shape of the answer, such as a list, table, or short paragraphs. Constraints: the rules, such as length, tone, and what not to do. You do not need all five for a simple question, but for any repeatable task, all five belong in the prompt.

Role: You are an editor for a small business blog.
Task: Turn the notes below into a blog outline.
Context: Audience is small shop owners. They want practical steps, not theory.
Format: A title, then five section headings, each with two bullet points.
Constraints: Use only the facts in the notes. No invented statistics. Keep headings under eight words.
Notes: [paste your notes here]

Write constraints as direct rules, not wishes. "Do not invent statistics" works better than "try to be accurate", because it names the behaviour to avoid. The same applies to guessing: state plainly what to do when information is missing, whether that is asking a question, leaving a blank, or flagging it.

Use examples to teach format and quality

Examples teach faster than descriptions. If you want a particular format, show one example of it. If you want a particular tone, show a paragraph in that tone and one to avoid. One good example usually beats a paragraph of adjectives, because adjectives mean different things to different readers, while an example is concrete.

Keep examples short and on-topic. Two examples that match the task are better than five that partly match it, and examples should use the same format you ask for, so the model can copy the shape, not just the idea.

Break large tasks into a sequence

Large tasks fail as single prompts because they ask for too many decisions at once. Break them into steps, each with its own prompt and its own check: first an outline, then a draft of one section, then a review of that section, then the next. Check each step before moving on, because an error in step one propagates through every step after it.

When a step's output will feed the next prompt, keep it in a consistent format, such as a list or a table, so the next prompt can use it without re-interpretation. Chains work when each link is clean; they fail when an early link is vague and later links build on the vagueness.

Test, compare, and keep what works

Testing means running a prompt on a small set of real inputs, including difficult ones, and recording what happened: where it worked, where it failed, and what kind of failure it was. Compare versions by running them on the same inputs, not on different ones, or the comparison tells you nothing.

Keep the winning prompt, with a note about what it is for and what inputs it was tested on. A prompt library built this way, from tested prompts rather than collected ones, becomes a real asset: every entry is a small, proven tool, not a guess.

Frequently asked questions