PTN—19
Stage and Conversation
Split the interface into a persistent workspace for the work and an adjacent chat for directing it.
Chat is a great steering wheel and a lousy workbench. The stage-and-conversation layout accepts both facts: the work — a document, a codebase, a design, a spreadsheet — holds the primary surface with full direct-manipulation editing, while the conversation runs alongside as the channel for instructions, questions, and narration. Neither pretends to be the other, and the person moves between talking about the work and just doing it.
The pattern’s real product is shared context. Because the stage is visible to both parties, the conversation can be deictic: “make this section shorter” resolves against the current selection instead of a paragraph of description. Every AI-native editor that works — Cursor’s editor-plus-chat, canvas modes in the major assistants — earns its fluency from this shared frame. Chat-only tools force people to describe what a pointer could simply indicate.
The design risk is divided attention. Two live regions compete for focus, and changes appearing on the stage while the user reads the chat can go unnoticed or feel unauthorized. The stage must signal what the AI changed — highlights, diff marks, pending states — and the conversation must reference stage locations concretely. When the two panes stop pointing at each other, the layout decays into two apps taped together.
When to use
- Sustained work on a persistent object — code, documents, designs — where chat directs rather than contains.
- When people need direct manipulation and AI assistance on the same object without switching tools.
- When “this” beats description — selection and pointing are the natural way to reference the work.
- Sessions long enough that the transcript alone would bury the current state of the work.
Design considerations
- 01Give the stage the dominant share of the viewport; the conversation is an instrument panel, not a co-star.
- 02Mark every AI-made change on the stage — highlights or diff coloring — until the person has seen it.
- 03Feed the current selection into the conversation’s context automatically, so “this” resolves without pasting.
- 04Let chat messages link to stage locations, and scroll the stage there on click.
- 05Support full manual editing on the stage; falling back to prompts for a typo fix is a pattern failure.
- 06Reconcile concurrent edits explicitly — never let an AI write-back silently clobber a human keystroke.
- 07Let the conversation collapse entirely when the person is heads-down in the work.
Pitfalls
- ✕AI edits landing silently on the stage while the user reads the chat, discovered only much later as mystery changes.
- ✕A chat pane blind to the stage, forcing people to paste the document into the conversation about the document.
- ✕Stolen focus — the stage grabbing scroll or cursor position whenever the assistant writes.
- ✕Two panes narrating independently, doubling the reading load instead of splitting the labor.
- ✕A read-only stage that demands a prompt roundtrip for edits the person could type in two seconds.
In the wild
- Cursor
- The editor is the stage and the chat pane directs it — the agent’s edits land as reviewable diffs in the code, referenced from the conversation.
- ChatGPT
- Canvas puts the document or code in an editable primary pane with the chat alongside; both the person and the model edit the same surface.
- Claude
- The artifact panel functions as a stage: the conversation revises the object in place while the person reads and copies from it directly.
- Figma
- AI features operate on the canvas itself — generated and edited layers appear in the working file, with the prompt UI subordinate to the design surface.