Expansive in the thinking lane, disciplined in the commit lane
Ask an AI to think with you and you want width: pull the internet, challenge the premise, propose a structure you hadn't considered. Ask it to apply the change you agreed on and you want a scalpel: the edit touches one slide, and nothing else. These are opposite virtues — and most tools run both through one dial.
That one dial is why the category's bug reports all sound the same: "I asked it to fix the title and it redesigned the deck." Or the mirror image: "I asked it to brainstorm and it timidly restated my own idea back to me." Turn the dial up and your edits get exuberant. Turn it down and your brainstorms get beige. The model's eagerness is a single temperament applied to two jobs that need opposite temperaments.
Two lanes, structurally separate
In Doric, thinking and committing are different lanes with different rules — not different prompts, different machinery.
In the thinking lane, Phoebe is deliberately expansive. She brainstorms wide, researches on her own initiative rather than waiting to be told, pulls in what the world already knows about your problem, and proposes structure — breaking your coarse idea into the finer pieces you'd have found eventually, and a few you wouldn't have. Nothing in this lane mutates an artifact. The conversation is allowed to be wrong, half-formed, contradictory, ambitious. That's what conversations are for.
The commit lane is the opposite temperament, enforced by the write path itself. A change lands as a proposal — marked, toggleable, accepted piece by piece — never as a silent mutation of your work. The write is scoped before it happens: this edit claims this slide, this file, this node, and the runtime holds it to that claim. Then a deterministic layer commits the change and reads it back to confirm the artifact says what the proposal said. The edit touches one slide. Nothing else moved, and the system can prove it.
Creativity is a conversation property. Discipline is a write-path property. Keep them in separate lanes and you never have to trade one for the other.
Why prompting can't fix this
Every tool in the category has tried the prompt-level version: "be creative when brainstorming, be conservative when editing." It decays. Over a long thread the instruction fades, and the model's underlying eagerness — whatever it is — leaks back into whichever job it's doing. You cannot make a behavior reliable by asking for it once.
The reliable version puts the two temperaments in places a model can't blur. Expansiveness lives in the agent, where it belongs. Discipline lives in the runtime — scoped writes, proposals over mutations, commit-and-read-back — where no amount of generative enthusiasm can reach it. The model is free to be imaginative precisely because the write path is not.
The mechanism, in one list: research-first brainstorming with no write access · every change rendered as a reviewable proposal · scoped writes with a claimed blast radius · a deterministic committer that reads back what it wrote. Wide where wide helps you; narrow where narrow protects you.
What it feels like
You feel it as trust moving in both directions. You can hand Phoebe a half-idea without bracing for her to overwrite your work, and you can accept her edit without diffing the whole document afterwards, because "the whole document" was never in play. The wildest conversation in the room and the most conservative edit on the artifact — at the same time, from the same colleague.