Prompt Engineering

Prompt engineering is writing the instructions you give an AI model so it does what you actually wanted.

The name oversells it slightly. It isn't engineering in the bridge-building sense — it's much closer to briefing a freelancer. A vague brief gets you something vaguely right. A brief that says who the audience is, what good looks like, what to avoid, and shows an example of a job done well gets you something usable. Models respond to the same things people do.

Why it matters more than it sounds

The gap between a lazy prompt and a considered one is not a small quality bump — it's often the difference between a feature that ships and one that gets quietly abandoned. The same model, on the same task, can be unusable or genuinely good depending entirely on how the instruction is framed.

That has a direct commercial consequence: most "the AI isn't good enough" conclusions are actually prompt problems, and they get diagnosed as model problems. Companies switch models, or abandon a use case, when the fix was in the brief.

What actually works

• Be specific about the output. "Summarise this" invites anything. "Summarise this in three bullets for a non-technical director, leading with the financial impact" gets you a usable answer.
• Show, don't just tell. One or two worked examples beat a paragraph of description.
• Give it a role. "You are a technical reviewer checking for security issues" narrows the space meaningfully.
• Say what not to do. Explicit constraints — don't speculate, say "I don't know" if the answer isn't in the source — reduce made-up answers substantially.
• Break up big jobs. Several focused prompts beat one that tries to do everything.

The limits worth knowing

Prompt engineering cannot make a model know something it wasn't trained on and isn't told. If the answer lives in your internal documents, no phrasing will retrieve it — you need retrieval-augmented generation to put the documents in front of the model.

It also can't fully fix reliability. A better prompt raises the odds of a good answer; it doesn't make the model deterministic. Anything that must be right every time needs a check outside the model — validation, a second pass, or a human.

And prompts are brittle across models. One tuned carefully for one model can degrade on another, which is a real maintenance cost people rarely budget for.

How we approach it

Prompt engineering is part of how we build AI features, not a service we sell on its own — it sits inside our AI solutions work. In practice it's iterative: draft, test against real cases, find where it breaks, tighten, repeat. The work is in the failures, not the happy path.