What Is Prompt Engineering? – Definition and How It Works
Learn what prompt engineering is, which techniques reliably help, what has been superseded by better models, and how to evaluate a prompt properly.
Prompt engineering is designing the input to a language model so its output is accurate, well-shaped, and consistent. It is closer to specification writing than to incantation.
1. What Actually Helps
| Technique | Why it works |
|---|---|
| State the task precisely | Ambiguity is resolved by guessing otherwise |
| Give the output format | A declared shape is followed; an implied one is not |
| Provide examples | Demonstration conveys conventions description cannot |
| Supply the context | Facts in the prompt beat facts in the weights |
| Permit refusal | Makes "I don't know" a likely output rather than an unlikely one |
| Constrain with a schema | Turns malformed output into a caught error |
Most poor results come from an underspecified task rather than from a suboptimal phrasing.
2. Structure Beats Cleverness
A workable prompt has recognisable parts: the role or context, the task, the input, the constraints, and the output format. Separating them explicitly — with headings or delimiters — outperforms a dense paragraph containing the same information.
Clear structure also makes prompts maintainable, which matters more over time than any single-run quality gain.
3. Examples Are the Strongest Lever
When output must follow conventions, showing two or three correct examples conveys more than a paragraph of description — including edge cases, tone, and how to handle missing values.
Choose examples that cover the variety the model will meet, and make sure they are consistent with each other. Contradictory examples teach ambiguity.
4. What Has Been Superseded
- Elaborate personas. Marginal, except where genuine domain framing helps.
- "Think step by step." Reasoning models do this internally; explicit instruction adds little.
- Politeness and pressure. Threats and rewards are folklore.
- Delimiter rituals. Useful for clarity, not magic.
Techniques that mattered on earlier models often no longer do, which is why prompt advice ages quickly.
5. Where Prompting Stops
Prompting cannot install knowledge the model lacks, keep information current, or verify a fact. Those need retrieval, tools, and validation.
Recognising the boundary saves considerable effort: a system that hallucinates because the fact is not in its context will not be fixed by a better instruction.
6. Evaluate, Do Not Eyeball
A prompt change that looks better on one example is not a measurement.
- Build a set of inputs with known good outputs.
- Run the current prompt and record results.
- Change one thing.
- Re-run and compare on the same set.
- Keep the change only if it improves the aggregate.
Model outputs vary between runs, so single-example comparisons are frequently noise.
7. Maintaining Prompts
Version them alongside code, keep the evaluation set with them, and re-run it when the model version changes. A prompt tuned for one model release can regress on the next, and without a set to run, the regression is discovered by users.
8. Summary
Prompt engineering is specification: state the task, supply the context, declare the output shape, show examples, and permit refusal.
Its limits are the model's — no phrasing supplies missing knowledge or current data. Evaluate changes against a fixed set rather than an impression, and expect the technique list to keep shrinking as models improve.