Prompt engineering

A prompt is everything the model reads before it answers: the system instructions, the user’s request, any documents or data added for context, and examples of what good looks like. Prompt engineering is the work of getting that package right and keeping it right. OpenAI describes it as writing instructions so a model consistently generates content that meets your requirements.

The craft has moved on from clever phrasing. Current guidance from OpenAI and Anthropic focuses on precise instructions, a few input and output examples, clear roles for system and user messages, and above all evaluation: define what success looks like, test against real cases and re-run the tests whenever the prompt or the model changes. Anthropic’s docs also note that not every problem is a prompt problem; cost or speed, for example, can be easier to fix by choosing a different model.

Why it matters for a SaaS company

Prompts now sit inside products and internal workflows: the assistant in your app, the agent that drafts landing pages, the script that summarises sales calls. A prompt is effectively a spec that runs every time. When it is vague, output drifts off brand, formats break and edge cases slip through, and nobody notices until a customer does. Treating prompts like product copy or code, with owners, versions and tests, is what makes AI features dependable.

How DesignMe handles it

We treat prompts as part of the system an agent works inside, not the whole of it. When we set up AI workflows, the foundations step produces component and section rules, the CMS content model, a brand and tone guide and a set of example requests, which together give the agent the context a short prompt cannot. In the pilot we run the workflows on real requests from the team, check drafts against their standards and fix what the agent gets wrong, then hand over a known-limits list so everyone knows where it needs a human. A clear brand identity helps here too: tone and visual rules written down for people work for agents as well.

Example

Say a SaaS team uses an AI agent to draft customer stories, and every draft opens with “In today’s competitive landscape”. Rewriting the prompt with the brand tone guide, two approved stories as examples, a banned-phrases list and a fixed outline (challenge, approach, result, quote) fixes the opening. Ten past interviews become a test set, and the team checks every prompt change against them.

Common mistakes

• Endless rewording. If output is wrong, check the context and examples first, not just the phrasing.
• No test cases. Without a fixed set of real inputs, you cannot tell whether a change helped.
• One giant prompt. Long, contradictory instruction lists confuse models. Split tasks and keep rules short.
• Ignoring model updates. A new model version can change behaviour. Re-run your tests when you switch.