multimodal-orchestration

Multimodal Orchestration

AI interactions increasingly span multiple modalities — text, images, voice, code, tools, structured data. Designing how these modalities work together is orchestration. The risk on every team: switching modality because it's available, not because it serves the user.

The point of multimodal isn't to use every modality. It's to use the right one at the right moment, and to handle the seams.

Modality selection

Each modality has strengths:

• Text: precise, editable, referenceable. Best for instructions, explanations, and nuanced content.
• Image: spatial, holistic, immediate. Best for layouts, diagrams, and visual concepts.
• Voice: natural, hands-free, emotional. Best for conversational flow and accessibility.
• Code: executable, precise, verifiable. Best for technical specifications and automation.
• Structured data: tables, forms, JSON. Best for comparison, configuration, and data entry.
• Tool use: actions in external systems. Best for execution, not generation.

The designer decides which modality the AI uses for each part of a response — and which modality the user is invited to use for input.

Cross-modal transitions

When the interaction switches modalities, design the transition:

• Text → image: "Here's what that layout could look like." AI generates a visual from a text description.
• Image → text: User uploads a screenshot; the AI describes or critiques it.
• Text → tool: AI writes a plan, then executes it.
• Voice → text: Spoken conversation captured and summarised as structured notes.

Transitions should feel seamless. The user shouldn't have to manually switch modes.

Modality conflicts

Sometimes modalities compete:

• Text says one thing, the image shows another
• Voice tone contradicts text content
• Structured output doesn't match the conversational frame

Establish a primary modality per interaction context. Other modalities are supporting — they elaborate the primary, never override it.

Decision rules

• Pick the modality the user can act on, not the one the AI can produce. A chart impresses if the user wanted a chart; if they wanted a number, the chart is overhead.
• Always preserve a text trace. Voice and image outputs are inaccessible to screen readers, unsearchable, and lost on session end. Even when text isn't primary, it's the audit log.
• When modalities disagree, primary wins — and flag the conflict. "The diagram shows three steps; the text mentions four — let me reconcile."
• Modality switches need announcement OR trivial reversibility. Either the user knows the mode is changing ("Here's a chart of that"), or they can dismiss the new modality in one tap.
• Don't switch modalities mid-utterance. Voice → image → voice in one turn breaks the user's processing rhythm. Group modality outputs at turn boundaries.
• The input modality belongs to the user. Never force the user out of the modality they chose to start with — the AI can output multimodally, but the user dictates input.

Anti-patterns

• Modality showcase: switching to a flashy modality (chart, voice, generated UI) because the system can, not because the user needs. The user sees the trick, not the help.
• Lossy translation: converting between modalities loses information that mattered. Voice → text drops emphasis and pause; image → text drops spatial relationships. Capture what was lost or keep both.
• Modality lock-in: the AI replies in modality X and the user can only respond in modality X. A voice question demands a voice answer, even when the user is now in a meeting.
• Unannounced switch: modality changes mid-interaction without warning. Breaks accessibility, breaks user expectations, breaks screen-reader users completely.
• Crossed primaries: two modalities both present themselves as authoritative. The user doesn't know which to trust when they disagree.
• Modality theatre: structured output (JSON, table) used for content the user wanted as prose. Looks rigorous; reads worse.

When not to use this

• Single-modality products by design — search bars, code completers, transcription tools. Don't add modalities just because the model supports them.
• Constrained-input contexts (driving, medical-glove use, low-bandwidth) — pick the one modality that works in that context and commit; orchestration overhead is dead weight.
• Early prototypes — get one modality working well before adding seams. Most multimodal failures are unreadiness in the primary modality.

See also

• generative-ui — when the supporting modality is generated UI, this skill defines when to switch and generative-ui defines what to switch to.
• conversation-patterns — modality transitions are repair sequences in disguise; design them as such.
• progressive-disclosure — secondary modalities are often advanced features — disclose deliberately.
• feedback-loops — track which modality switches the user reverses; that's the loudest implicit-feedback signal you'll get.

Design Artefacts

• Modality maps showing which modality is used at each interaction point
• Cross-modal transition specifications (trigger, content, fallback if modality unavailable)
• Input/output modality matrices per feature
• Fallback definitions when a modality is unavailable (slow connection, screen-reader, accessibility settings)

Worked example — modality matrix for a design-review feature:

| Stage | AI output (primary) | AI output (supporting) | User input | Why | | --- | --- | --- | --- | --- | | Upload feedback | text confirmation | — | image (mockup) | User in image-thinking mode; AI confirms in text for a record. | | Critique | text (numbered) | annotated image overlay | text | Numbered text is referenceable; the overlay locates each point. | | Suggest fix | annotated image | text rationale | text or click on point | User picks: visual fix (one tap) or written justification. | | Iterate | text | image diff | image (revised mockup) | Iteration cycle — text dominates because the user wants comparison, not impression. |

Note the pattern: the AI uses the modality that matches the user's input or the modality that's most actionable, with text as the backbone. Image is supporting most of the time, primary only at the precise moment a visual change is being agreed.

Adapted from work on multimodal interfaces and the principle that modalities should map to the user's representation of the task, not the system's capabilities (Maybury & Wahlster on intelligent multimedia interfaces).