Design an exhibition campaign by voice

“Bring the flower much closer. I want it to almost spill out of the poster.”

A red flower grows beyond the frame. “Make AFTER HOURS enormous.” Two lines of type cross its stem. The date gets a sharp accent; the practical details settle into an aligned group. Then the same identity becomes a square for social media and a mobile ticket page.

The designer keeps talking about the work. The canvas keeps changing beneath that conversation.

Near the end, the direction changes: cream paper, a contained image, something more like a gallery invitation. The designer sees it, prefers the earlier night version and asks to bring it back—keeping the social and ticket work. That ability to explore while preserving what already works is central to the experience.

The experience

The film follows all eleven changes to After Hours, a fictional exhibition campaign, in one continuous session. Each result becomes the next direction’s starting document. The active format fills the workspace; the other formats remain visible alongside it. Before trying the cream alternative, the designer selects the poster explicitly. At the end, all three formats fill the workspace together. The flower is an original generated image. Type, geometry and campaign content remain separate editable objects.

The session moves through several kinds of design decision:

• Framing and hierarchy: enlarge the image, then give the event title enough scale to become part of the composition.
• Selective emphasis: color the date while preserving everything else; align the venue beneath it.
• Adaptation: create a square composition with its own crop, then a ticket page with a clear action. The formats share event details, without sharing every layout decision.
• A format-specific correction: give the square a prominent date band; give the ticket page more room around its admission details.
• An alternative and a return: try a quieter poster treatment, then restore the exact preceding poster while keeping the other formats.

How fast was the recorded session?

All eleven spoken requests selected their intended operations. The visible updates followed speech by approximately 2.08–2.24 seconds. The film runs at normal speed with no cuts inside the session. The pauses and result holds are part of the recording.

The model’s HTTP requests took 227–306 ms. Local CPU speech recognition took 1.41–1.49 seconds, accounting for most of the interaction delay. These measurements describe different parts of the loop: fast model decisions alone do not make the entire spoken interaction subsecond. Reducing the recognition delay is a concrete next improvement.

The inputs are generated speech recordings, played in full through actual speech recognition and a prebuilt fast model. End-to-end timing includes recognition, the model request and application of the result. We estimate the acoustic endpoint from the last 10 ms audio frame above −45 dBFS, then compare it with the browser’s audio-output clock and next paint opportunity. Audio/video alignment is bounded to approximately 43 ms. This is a measured prerecorded-input session with warmed recognition, rather than a measurement of a hardware microphone, display or unfamiliar requests. Captions follow the original spoken words; the interface retains the actual recognition text, including its imperfect transcription of “AFTER HOURS.”

The practical ambition is a short enough loop to examine a change while the next direction is still forming. This recording makes both the working interaction and its current recognition delay available to inspect.

The model and the tool

“Put the type in front of the stem” has an executable meaning only if the system knows how this document is drawn. The model receives the painting order, headline geometry, protected event details, current selection and actual document state. It can distinguish putting text above an image in the layer stack from moving it upward on the page.

The request produces three answers in one model call:

• Scope: which format is being changed or inspected?
• Family: is this framing, hierarchy, arrangement, adaptation, atmosphere or another supported operation?
• Operation: which available transaction actually implements the whole direction?

These questions share the same document. None is given another question’s predicted answer. A deterministic check then requires the three answers to agree before applying a transaction.

Each transaction declares its effects and preconditions. Enlarging the title changes its size and baselines; the title remains text. The square receives an independent composition. The cream treatment changes the poster’s background, type and image frame together. A request outside the available vocabulary leaves the document unchanged.

The tool also checks the state it is about to change:

• The returned operation must belong to the declared contract.
• The document revision must still match the one supplied for interpretation.
• A format must exist before an operation can rearrange it.
• The resulting scene must satisfy its geometry and content constraints.
• Restoring the night version requires the exact preceding state, including the social and ticket work the instruction preserves.

There is a useful division here. The model interprets a direction in context. The tool determines what the operation does. The designer judges the result on the canvas.

The representation has to describe the real work

Development exposed a small but consequential mismatch: an alignment operation claimed to share the headline’s left edge, while the renderer put those elements at different coordinates. The implementation now computes the geometry once for both rendering and model context. “Aligned” refers to the same relationship in both places.

That example matters beyond layout. A model needs an accurate description of the tool it is using. Current state, operation definitions and rendered consequences must agree. The recorded session followed a separate fifteen-case check: eleven supported directions and four refusals. These are authored development examples, not an independent estimate of general design ability.

Research: keep alternatives available

Terry and Mynatt’s Side Views begins with expert image-editing practice: people experiment, branch, compare and iterate, while an interface often exposes only one current document. Their prototypes offered persistent previews that updated with the active document, including composed operations and ranges of parameter settings. A preview could remain available without changing the working source.

That is a productive design connection for a conversational canvas. “Try cream paper” can become a view to compare, while the accepted campaign stays available. The paper presents an interface mechanism and implementation; it leaves formal user testing as subsequent work. Our current canvas records an applied alternative and exact undo. Persistent, automatically refreshed alternatives are a further capability to develop.

There is also empirical reason to examine how alternatives are organized. In Dow and colleagues’ parallel-prototyping study, 33 participants created web advertisements with equal prototype counts and allocated time. One group received feedback after each prototype; the other made several alternatives before feedback. The parallel condition produced more varied work and stronger results on the study’s performance measures, including independent ratings. The intervention changed the structure of making and critique; it did not test voice input, AI generation or subsecond interaction.

The useful question for this tool is therefore more specific than whether it can make edits quickly: can it make a consequential comparison easy enough to remain part of the work? A poster’s restrained version and its theatrical version might remain visible together. A designer could keep the second composition and borrow only the first one’s treatment of practical details. That requires explicit, composable design operations, plus a way to show what a combination changes.

Hutchins, Hollan and Norman’s Direct Manipulation Interfaces provides another distinction: reducing the distance between intention and an available operation is different from making the represented object feel directly available to act upon. Language can help with the former. An inspectable, editable canvas supports the latter. A good interaction needs both, including the moment when looking at the result changes what the designer wants.

Building it with SYSTEM1

The capability to develop is a model and a meaningful set of operations together. A campaign tool needs more than an instruction to “make good design.” It needs to know what can change, which relationships matter, what content must survive and how a person examines the result.

This prototype uses a prebuilt fast model with an authored operation vocabulary. A specialized model could learn that vocabulary from examples tied to the exact tooling revision, including refusals, partial requests and corrections in context. Current copy, geometry and accepted history would still be supplied at use time. Training would need evaluation on unfamiliar language and altered documents before claiming broader coverage.

An embeddable design surface brings those operations into the place where the work happens. Its callable contract needs to carry the request and document state, return a validated edit or a request for clarification, and preserve the conditions under which the answer is usable. A visual surface can then expose the same operations through speech, selection and direct controls. Building semantic tools explains how professional concepts become operations that models can learn.

The company mission is to streamline the development of this whole capability: discovering useful operations with the business, designing their contracts, creating and evaluating training data, and delivering the model with its tool and interface. A venue, publisher or retailer should be able to commission behavior around its own work without first assembling that production process.

Where this could lead

Consider a publisher preparing a touring exhibition. Its materials include street posters, venue screens, invitations and a programme. Some decisions belong to the whole identity; others belong to a viewing distance, language or occasion. A realtime tool could make those relationships available to direct.

• A direction that remains active. “Keep the work’s title dominant. Make the date legible at a glance. Let the image crop before shrinking the title.” After the designer confirms the intended relationships, each format can respond to later copy changes. Independent checks examine text fit and available space; a shared check preserves the campaign’s agreed hierarchy. A format that cannot satisfy the rules presents the conflict on that artboard.

• Alternatives with a purpose. “Show a version for someone passing on a bus, and one for someone holding the invitation.” The tool develops different reading sequences around the same exhibition. The designer can compare what each gives prominence to, then change that intention. Legibility and reading order need their own tests; a label such as “bus version” does not establish either.

• A comparison that teaches the next request. Keep an energetic composition and a restrained one side by side. “Keep the first one’s scale, but use the second one’s spacing around the date.” A model resolves which features the person means; the tool checks whether they can be composed. The result may reveal a direction the designer could not have specified before seeing the alternatives.

• A spatial rehearsal. Carry the same identity into signs for the venue. “At this entrance, help people find the upstairs gallery before introducing the event.” The relevant operations now concern sightlines, reading sequence and placement. Geometry checks physical constraints; contextual decisions identify which information belongs at each point in the journey. The designer can walk through a simulation and refine the sequence while seeing its consequences.

These are proposed extensions. They suggest a larger role for fast decision models: maintaining a working relationship between a person’s intention and a changing set of objects. A direction can remain inspectable, revisable and available across the next piece of work. The designer gains more room to ask what the work could become—and a concrete way to find out.