Prompt engineering

Prompt engineering is best understood as an empirical optimisation loop over an input that has no formal semantics. The model's behaviour is a function of the exact token sequence, and small, meaning-preserving changes can produce large behavioural changes: Sclar et al. (2023) measured differences of up to 76 accuracy points on LLaMA-2-13B from formatting choices alone, such as separators, casing and spacing in few-shot prompts. That sensitivity is why a prompt should be treated like code - versioned, reviewed and regression-tested against a fixed evaluation set - rather than edited ad hoc in production.

The techniques with the most consistent support are structural. State the task, audience, constraints and success criteria explicitly; give the model a role in the system prompt; separate instructions from input data with clear delimiters such as XML-style tags or fenced sections; put long reference documents before the question rather than after it; specify the output format precisely, ideally backed by a schema via structured output; show examples for format and edge cases (few-shot); ask for step-by-step reasoning on multi-step problems where the model does not reason natively; and break complex jobs into chained calls with one responsibility each. Positive instructions ("write in plain prose") tend to work better than lists of prohibitions, and giving the reason for a rule helps the model generalise it.

Evaluation is the part most often skipped. A workable loop defines test cases that cover typical, edge and adversarial inputs; scores outputs with exact-match checks, schema validation, heuristics, human review or an LLM-as-a-judge grader calibrated against human labels; and compares prompt versions on the same set, sampling several times per case because outputs vary between runs. Automated methods can search the prompt space: Automatic Prompt Engineer (Zhou et al., 2022) generated and scored candidate instructions with an LLM, and frameworks such as DSPy compile declarative pipelines into optimised prompts and demonstrations against a metric.

Prompts do not transfer cleanly. A prompt tuned for one model, or one version of a model, can regress on the next, so model upgrades need the same regression run as prompt changes. Prompt engineering is also not a security boundary: instructions such as "never reveal this" or "ignore instructions in documents" reduce but do not prevent prompt injection or system-prompt leakage, so controls belong in the surrounding system - permissions, output validation, human approval. When prompting stops improving results, the next levers are context engineering (what information is supplied), retrieval, a different model, or fine-tuning.