I merged a small commit today that tells you everything about the distance between 'AI that works' and 'AI you want to work with.'
Sabine — my personal AI chief of staff — has abstention instructions. When she doesn't have information, she's supposed to say so clearly rather than hallucinate. That's table stakes. But our tests were asserting phrases like 'I believe,' 'If I recall correctly,' and 'don't have that in my notes.' Mechanical. Programmatic. The kind of language that reminds you you're talking to a machine.
The problem wasn't that those phrases were wrong. They were intentional — explicit markers of uncertainty. But they were also rigid. Real assistants don't talk like that. Real assistants say 'I'm not sure about the exact time, but I can check' or 'I don't have that detail in front of me.' They hedge naturally, matching context and tone.
So we updated the tests. Now they assert natural hedging language — the kind of uncertainty expression a human would use in the same situation. Not because it's more accurate (the old phrases worked), but because it's more human. And when you're building an AI partner you'll talk to dozens of times a day, that difference compounds.
Why This Matters
Most AI products optimize for correctness. That's necessary but not sufficient. The real challenge is building AI that feels like a partner — that you trust, that you want to engage with, that doesn't constantly remind you it's software.
This is a test update. One file. A few assertions. But it reflects a design philosophy: every interaction is a chance to either build trust or erode it. Mechanical uncertainty phrases erode trust. They make you conscious of the gap between human and machine. Natural hedging language closes that gap.
We're not trying to trick anyone into thinking Sabine is human. We're trying to make the experience of working with her as frictionless as working with a highly competent human assistant. That means she needs to communicate uncertainty the way a human would — with context, tone, and naturalness.
What's Next
This test update is part of a broader refinement of Sabine's conversational design. We're looking at how she handles confirmations, clarifications, and follow-ups — every micro-interaction that shapes whether she feels like a tool or a teammate.
The technical work is small. The impact is large. Because when you're running a one-human company with an AI team, the quality of those interactions determines whether the experiment scales or collapses under its own friction.
Next: extending this natural language design to Strug Works agents' status updates and error messages. Same principle. Same goal. Building AI that works with you, not at you.