Prompt engineering
A model does what its instructions say, which is rarely quite what you meant. We write the instructions, test them against real cases, and keep testing as the models change.
Start a projectWhat is included
System prompts
The standing instructions behind an assistant: what it is for, what it knows, how it speaks and what it must refuse.
Test sets
A collection of real questions with known good answers, run every time the prompt or the model changes.
Guardrails
Limits on what the model may say or do, and checks on its output before a customer sees it.
Grounding in your material
The model answers from your documents and shows where the answer came from, or says it does not know.
Prompts for your team
A written set of prompts and working habits for the AI tools your staff already use.
Model comparisons
The same task run on Claude, Gemini, Grok, open-source models and GPT, so you can choose on results and cost.
How we go about it
- 01
Define a good answer
We collect real examples and agree what a right answer looks like. Without that, nobody can say whether a change made things better.
- 02
Write, then test
We draft the instructions and run them against the examples, then change one thing at a time.
- 03
Try to break it
Odd questions, angry questions, and text written to trick the model into ignoring its instructions.
- 04
Keep it under version control
Prompts are stored, reviewed and rolled back like code, so a change that makes things worse can be undone.
A good fit when
- Your assistant gives a different answer every time.
- The output needs heavy editing before anyone can use it.
- A model update changed how your tool behaves and nobody knows why.
Prompt engineering questions we get
Writing and testing the instructions given to an AI model so that it produces accurate, consistent output. It covers the wording, the examples, the reference material the model is given and the checks on what comes back.
For one answer, no. For a tool that hundreds of customers will use, the instructions have to hold up against questions nobody predicted, and that takes testing.
No. Each model responds differently to the same instructions. We test on the model you will use, and test again when it is updated.
An attack where text in a message, a document or a web page tells the model to ignore its instructions. We design for it: the model gets only the access it needs, and its output is checked before anything acts on it.
Often paired with
Need help with prompt engineering?
Tell us what you have and what you need. A sentence or two is enough to start.
Start a project