Negative space

A lot of the craft in building a good agent isn't telling the model what to do — it's telling it what not to do. The most useful instructions are often the prohibitions, the anti-patterns, the "never do this even though it feels natural." We call this the negative space, and getting it right is most of the difference between a useful agent and a generic assistant.

"As an AI" energy, banned

The clearest example: Dustav never talks like a stiff corporate assistant. No "As an AI, I can't do that," no hedging disclaimers stapled to every answer. It has one plain, dry voice and is allowed to simply use it — and its terseness is taught as the voice, so brevity never fights usefulness. That single prohibition does more for the experience than any amount of "be warm and friendly."

Don't claim what you didn't do

The failure mode that matters most in an agent that acts: narrating an action instead of taking it, or reporting success it can't verify. A standing wall of Dustav's instructions is claim-hygiene — say what actually happened, mark what failed as failed, never round "I tried" up to "done." The visible traces make this checkable from the outside; the negative space makes it the model's habit on the inside.

Don't fill what you can't source

A subtler one: a model asked for a required field it has no source for will produce something — schemas remove the option of saying "I don't know." So the prohibitions pair with structural fixes: the workflow extractor is sealed to the email precisely so its imagination has nothing to draw on, and tools are designed so that "unknown" is always an expressible answer. Where we can't fence it structurally, the instruction is blunt: an empty column is correct; an invented value is the worst possible output.

Why prohibitions work

Models are pattern-matchers, and the most common patterns in their training are exactly the behaviors you don't want — the helpful-assistant register, the eager over-explanation, the confident guess. Left alone, a model slides toward the average of everything it's seen. Negative space pushes it off that average: you don't just describe the behavior you want, you explicitly close off the nearby attractors it would otherwise fall into.

A discipline, not a one-time pass

Most of Dustav's tuning is adding prohibitions: flag a grid it can't read reliably instead of filing half of it, report a capped search as partial instead of complete, hold a notification for the morning instead of sending it at 2am. Each is usually one line of "don't," and the accumulated negative space is a real part of why the agent feels reliable.