AI brainstorming on a canvas
Brainstorming dies in a linear chat. The second idea overwrites the first, the model anchors on whatever you said last, and by the tenth message you are scrolling to remember what you liked. On a canvas, every direction you try stays on screen, side by side, until you decide.
fork
Take any promising answer and fork a branch per direction: one for the bold version, one for the safe version, one for the weird version. The main thread stays clean and every branch keeps the full story that led there.
reference
Writing the brief for direction B and want the constraint you settled in direction A? Reference that exact message with @ instead of re-explaining it.
merge
When two directions both have something right, ask one question with both branches in scope. You get a single answer that accounts for everything you explored.
Start with the question, not the answer
Open a workspace and put the real question in your first message. Let the model give its first take; that becomes the trunk everything else grows from.
Fork per direction, not per message
Each genuinely different direction gets its own branch. Three branches you can see beat ten regenerations you cannot. Different branches can even use different models: one direction on Claude, another on GPT, and compare temperament.
Kill ideas visibly
Dead ends stay on the canvas. That is a feature. Seeing why an idea died stops you from wandering back into it next week, and a text note pinned beside it records the reason in your own words.
Merge the winners
Select the branches that survived and merge them into one draft. Ask for the plan, the pitch, or the spec. The answer draws on everything in scope, and you can trace exactly which thread contributed what.
Common questions
How is this better than regenerating answers in a chat?
Regeneration replaces; forking preserves. When you regenerate, the previous idea is gone unless you saved it somewhere. A fork keeps both alive, visibly, with their full context intact.
Can I use different models on different branches?
Yes. Every paid plan reaches every model, and the model is chosen per message. Running the same question through two models on two branches is a two-click experiment.
No credit card to start. Your first workspace is already waiting.