Prompt Engineering

What prompt engineering is

Prompt engineering is getting a model to do what you actually meant, through the words, structure, and examples you give it, not through a setting you flip. The same model, on the same task, produces a noticeably better or worse result depending on how the prompt is written.

Five levers that make a prompt work

• Explicit task: state exactly what you want, not what you’re avoiding. “Summarize this in three bullet points for a non-technical audience” beats “summarize this”
• System message: split system and user messages. The system message sets role, tone, and constraints once, while the user message carries the input
• Output format: state the output format you want. If you need JSON, a fixed length, or a specific structure, say so directly rather than implying it through example alone
• Few-shot examples: show, don’t just tell, for nuanced judgment. An assistant message is a valid part of a prompt, so one or two example responses written as assistant turns teach a style or a judgment call more reliably than a paragraph describing it
• Relevant context: only include context the model needs. More context isn’t automatically better; irrelevant material distracts the model and adds cost without adding accuracy

Here’s a weak prompt and the same task with all five levers applied:

Weak prompt:

Here's the customer's full message history, all 40 messages. Look at the most recent one and tell me how they feel about it.

Same task, with all five levers applied:

System message:

You are a support assistant. Be concise and avoid jargon.

User message (example turn):

Classify the sentiment of this customer feedback as Positive, Negative, or Neutral, and give a one-word reason.

Feedback: "The product arrived broken and support never replied."

Assistant message (example turn):

Negative - defect

User message (real input):

Feedback: "The app keeps crashing when I try to export my report."

From symptom to lever

When a prompt’s output goes wrong, it’s tempting to guess at the cause. It’s more useful to ask which lever above actually controls it, since most failures come down to exactly one of the five.

Symptom	Lever	Fix
Instruction gets partly ignored	Explicit task	State the instruction plainly instead of burying it in a longer prompt
Tone drifts across runs	System message	Define persona and tone in the system message instead of leaving them implicit
Format or length is inconsistent	Output format	Spell out the exact structure or length you want
A judgment call misses the mark	Few-shot examples	Add one or two assistant message examples showing the call you want
Model hallucinates or drifts off-topic	Relevant context	Give it only the context it needs, and tell it to say “I don’t know” when that’s not enough

The iteration loop

Prompt engineering rarely lands on the first try. Run the prompt against real inputs, evaluate the outputs instead of eyeballing them, and commit and compare versions to check whether a change helped. Change one thing at a time, so if the score moves, you know which change caused it.

Keep exploring