From 50+ AI trainings to the Gen-AI Labs: teams at Leiden University build internal prototypes in three sessions

At Leiden University it started with a broad movement to make staff AI literate. We delivered more than fifty AI foundation trainings for university staff. As a result, people recognised opportunities and risks more quickly, and could talk about them together without working at cross purposes.

The approach was open to staff from different departments: researchers, administrative teams, operational colleagues, people who keep processes running and people who create and verify new knowledge. That mix meant the use cases didn't stay theoretical, but were tested against reality straight away.

After the trainings came the logical next question: not just "how does it work", but "how do we put this to use so that it really helps us?". There was a clear need to go beyond inspiration, to work with real use cases, test ideas and build prototypes.

The challenge was to make the move from AI literacy to identifying, testing and further developing concrete, scalable solutions within the organisation.

That question gave rise to the Gen-AI Labs: a programme less about broadcasting and more about building. The university set up a safe online environment in which staff could work with language models. Not isolated experiments, but a controlled place where you can learn, test and improve as an organisation. In that environment teams turned ideas into prototypes by writing system prompts and testing iteratively.

The programme consisted of three lab sessions, each with a clear rhythm: sharpen what you build, build and test, and then present with a plan for next steps. We coached teams on grounding, reliability and evaluation: how to make sure answers are based on sources, how to write instructions for language models, and how to let systems work together across multiple steps. This led not only to better results, but also to more trust, because you can explain why an answer is correct and when you should not use it.

Teams also learned to build system prompts, with best practices around role, task, context, constraints, examples and a consistent tone of voice. Structured output was an important part, because many internal processes only really speed up when the output is predictable: think of tables, lists, category choices or formats that can flow straight into a next system. We also helped shape product features around the interaction, such as source references, feedback loops, versioning of prompts, fixed test questions and simple guardrails for users. Those choices make the difference between a nice demo and a prototype that is ready for further development.