Human-Centered Context Engineering: an interview with Thomas Immich

AI Human Centered Design Personas User Research LeanScope Lean UX UX
by Thomas Immich
Portrait of Thomas Immich on a blue background, framed by a white, sweeping ellipse. To the left, a compass drawn from a single white line with playful loops, its needle pointing at Thomas.

"Context engineering" is well on its way to becoming the next umbrella term of the AI world. Almost always, it refers to what you give an AI model about a codebase. Our Managing Director Thomas Immich puts "human-centered" in front of it and means something deeper: the entire context a team has about the mental models and needs of its users, in other words the "context of use".

For the project Basiswissen KI, Michael Überschär spoke with Thomas at length: about AI personas that are allowed to shoot down ideas, about bots with human godparents, about deliberately built-in friction and about the question of who actually owns the context of an AI agent. Here are the key ideas from the conversation.

Context that is not in the code

From a technical point of view, the context in context engineering is usually derived from code: a coding agent like Claude Code looks at the codebase and knows what to do next. Domain experts tend to rely on knowledge context instead, such as wikis, manuals or tickets. Human-centered context engineering combines both, with a specific focus:

"This is about the knowledge that has been built up about the person the product ultimately has to please: the user."

That is exactly the knowledge many companies lack. Domain knowledge usually exists, but rarely from the users' perspective and even more rarely aligned with their mental models. This is where LeanScope AI comes in, the product discovery platform that grew out of Centigrade's project business: it helps teams build up user research insights and make them usable for AI through knowledge graphs and retrieval. The claim behind it: "See the world through the eyes of your audience."

Bots with godparents: the in-house experiment

Alongside people, our team page also lists a few bots, such as T0-N1, our voice-and-tone bot, or E. Le Fand. It started shortly after ChatGPT 3.5 was released with P.F.U.X., a bot for Thomas' podcast at the time, "Prompts for UX". It quickly became clear that a well-crafted character does something to people. It is not just text, but a personality that delivers that text.

This also sparked a legitimate discussion: do people feel replaced when AI has not only impressive skills but its own personality, too? The answer was the godparent model. Every bot has a godparent, a person who trained it, maintains its prompt and stands up for it.

"The godparent is the godfather or the godmother. After that, nobody feels subordinate to the AI anymore."

This attitude is also reflected on our website: "People" is the first menu item, while AI comes further back. AI matters, but the enablers of complex, high-quality digital systems are the people on our team.

How do you know a persona is right?

AI personas do not replace real users. They are a compass pointing towards them. A persona never represents a single person, but the sum of the characteristics of similar people. That is why Thomas draws a clear line between the proto-persona, generated on demand from the AI's general world knowledge, and the verified persona, which is grounded in real interviews.

An example from a Centigrade project shows how big the difference can be: in its proto state, a mechanical engineer persona found the idea of logging in to the machine via facial recognition "innovative, awesome". After just three in-depth interviews with real mechanical engineers and uploading the transcripts, the same persona replied: "I wear a hard hat and goggles at the machine, I don't think that will work."

"Your own idea may sound cool, but the persona shoots it down mercilessly. That kind of thing can genuinely ground us humans."

To keep teams from developing against their own confirmation bias, LeanScope relies on assumption tracking, a principle from Jeff Gothelf's Lean UX: disputes in a workshop are not argued out but noted as assumptions that still need evidence. The persona improves with every new insight, and user researchers remain a permanent part of the process. Thomas sees the greatest risk in teams that have never done any user research, because they are the most likely to take everything the persona says at face value. That is also why the partnership with TestingTime is not a contradiction but a meaningful complement: TestingTime recruits real people for interviews, and LeanScope turns the results into lifelike AI personas.

Digital twins and the data protection trap

Thomas also clearly distinguishes personas from digital twins. A digital copy of an individual is a one-to-one relationship and therefore always carries a data protection trap.

"If I've hacked the digital object, I've more or less hacked the person. That's why, for us, personas are not one-to-one but explicitly one-to-many mappings."

Because a persona always represents the knowledge of many people, no statement can be traced back to a single individual. This avoids a personal data trap by design.

Where friction has to stay

Our industry likes to reduce friction, cognitive load and interaction costs. Thomas argues for deliberately building in friction in the right places. Anyone who sees every check as a brake risks "deskilling": we unlearn undivided attention, critical thinking and precise quality assessment.

That is why LeanScope is working on a "second voice": a second, critical AI that reviews the conversation with a persona and points out, for example, that a statement is not yet backed by any user research. A built-in nudge in the spirit of Jeff Gothelf's motto "get out of the building".

"We want AI personas to make people do more user research, not less."

Saddle or seat: mental models

Two people, one term, two mental models: when most people think of a bicycle, they think of a saddle. Someone who rides a recumbent bike thinks of a seat instead. Software is often designed with the mental model of UX professionals or, worse, of software engineers, and that clashes with the mental model of its users. Personas capture exactly these mental models in enough detail to be directly usable in user interface design. Approaches such as the Open Knowledge Format, which provides aliases and synonyms for terms, point in the right direction: knowledge is not absolute but depends on the context of use.

The shared campfire and the echo chamber

When software is increasingly tailored to individuals through personalized AI agents, what happens to the shared campfire? For Thomas, in the end it is not the tools but the stories we tell each other. Education, explaining how algorithms work and a conscious fight against our own convenience help against echo chambers, much like going for a run at five in the morning in the rain.

"If everyone always agrees with you, you are already in the echo chamber."

Looking ahead

How will Thomas know the journey has worked? Not by patting himself on the back, but by the sparkle in users' eyes and by a team that believes in what it builds. For him, success is a good match between values, purpose and the work reflected back by customers and users.

Read the full interview

The complete conversation, with many more questions and answers, is available in German on Basiswissen KI: Human-Centered Context Engineering – interview with Thomas Immich. If you want to dive deeper into persona agents, read our journal article AI Personas: Empathy-Amplifier or false pretences?

The quotes are taken from the interview "Human-Centered Context Engineering" by Michael Überschär on basiswissen-ki.de, published under the CC BY 4.0 license, and have been translated, shortened and summarized for this article.