Adobe · GenStudio for Performance Marketing

Horizon Canvas

A performance marketer can take what GenStudio generates and get it channel-ready without going back to an agency. That's where Horizon Canvas comes in, it's the part that handles everything after generation: reframing assets, extending them, cutting video, and launching the campaing.

Senior UX Designer Post-Generation Workflows Content Operations Interaction Systems Enterprise Platform 2025 – 2026

GenStudio content lifecycle, where Horizon Canvas fits: from first concept through adapting content to handing it off for activation

At a Glance

Company	Adobe
Product	Adobe GenStudio for Performance Marketing - Horizon Canvas
Users	Enterprise performance marketers: mid-senior marketing leads at Fortune 500 brands managing multi-million dollar paid media budgets across channels
Team	Multiple designers across Horizon Canvas. I led GenExpand and the entry point system, partnered on Reframe and Timeline, and led cross-surface consistency across the platform
Duration	2025 – 2026 · H1 2026 design cycle
My Role	Senior UX Designer: post-generation content operations, interaction systems, cross-surface consistency, collaboration with PM and engineering
Key Decision	Build Horizon Canvas around what happens after generation. Concept exploration, channel adaptation, and activation are one workflow, not three separate features
Outcome	A marketer can take an AI-generated concept, adapt it for every placement, and prep campaign-ready variants without sending it back to an agency

Why This Mattered

The Problem

Enterprise performance marketers were entirely dependent on creative agencies for any content adaptation. Every format change, every channel resize, every AI-generated asset that needed refinement required a specialist. GenStudio's promise was to give them a third dial — the ability to create and iterate on content themselves.

My Contribution

I shaped the post-generation content operations workflows that connect AI generation to campaign activation — leading GenExpand, establishing the interaction framework that unified multiple disconnected capabilities into one consistent experience, and partnering on Reframe and Timeline. My job was making five separately-built features behave like one product. GenExpand, Reframe, and Timeline each had different owners. I designed the interaction system they all shared.

The Outcome

A content adaptation platform where enterprise performance marketers move from AI-generated concept to activation-ready assets without leaving GenStudio — or routing back to a creative agency. The platform's capabilities (Expand, Reframe, Timeline) are the working parts. The transformation — from agency-dependent to marketer-led content operations — is the outcome.

The Work

AI-Powered Adaptation

Extend and adapt AI-generated content across formats, without going back to a designer

Unified Interaction Framework

One interaction pattern across every AI capability that scales automatically as the product grows

Concept to Campaign

End-to-end workflow: AI output → marketer edits → campaign-ready asset, no agency required

Outcomes

Horizon Canvas was in active development through H1 2026. Outcomes reflect delivery milestones, design handoffs, and architectural sign-offs, not post-launch metrics.

✓ Post-generation platform signed off at the January 2026 leadership review. Enterprise marketers can take an AI-generated concept to a campaign-ready asset inside GenStudio, without routing back to a creative agency.
✓ Every new AI capability ships with discoverability already handled. The Unified Interaction Framework means users don't pay a learning cost for features they haven't seen before.
✓ Marketers can extend, reframe, and activate AI-generated content on their own. GenExpand, Reframe, and Timeline give performance marketers direct control over content operations, without a specialist at every step.
✓ Enterprise performance marketers can prepare video campaigns without leaving GenStudio. Scenes, sequencing, and clip control now sit alongside image operations in one canvas, which wasn't possible before.
✓ Engineering built from complete specifications. Interaction states across all five AI capabilities were documented end to end, so development could run in parallel without design ambiguity.
✓ Marketers carry what they learn into every capability the platform ships next. One hover-triggered pattern surfaces all content operations, whatever the content type or AI feature.

Why Horizon Had to Change

The old experience was built for a world without AI.

GenStudio's creation surface was originally built on an HTML Canvas architecture. It worked for what it was designed for. But as Adobe moved GenStudio toward AI-native content creation at enterprise scale, the limitations compounded. The system wasn't wrong. It was designed for a different problem.

One asset at a time

The original experience was optimised for working on a single ad. That model broke down once performance marketers needed to generate, explore, and iterate across multiple concepts and dozens of format variants at once. There was no way to compare Concept 1 against Concept 3, or to see how one idea looked across Meta Stories, LinkedIn Banner, and Display side by side.

AI was bolted on, not built in

Early AI capabilities were added to an experience designed before they existed, so AI generation felt like a detour from the workflow rather than the centre of it. Editing after generation meant leaving the tool or going back to a creative specialist. The post-generation workflow, what happens after AI creates content, had no dedicated design at all.

No video architecture

Video was becoming a first-class content type for enterprise paid media, with Meta, Google, and LinkedIn all prioritising video placements. GenStudio had no workflow for managing it: no timeline, no scene management, no way to handle the temporal structure of a video ad. This wasn't a missing feature, it was a missing content paradigm.

Format adaptation required leaving the tool

Adapting an asset from 1:1 to 9:16 to 16:9 meant going back to a designer, opening Photoshop, or using a separate resizing tool. For a platform promising performance marketers creative independence, that was a fundamental gap. Every format change broke the workflow and re-introduced the exact agency dependency GenStudio was supposed to eliminate.

No scalable content operations foundation

As AI capabilities grew with generative expand, format adaptation, AI video there was no shared platform foundation to build on. Each new capability would need its own entry point, its own UI paradigm, its own discovery pattern. Without a foundation, the product would fragment into disconnected AI tools instead of cohering into a content operations platform.

Marketers still needed specialists for everything

Despite the promise of AI, performance marketers were still routing every creative decision through agencies and designers. The capability existed but the pipeline didn't. GenStudio needed a creation and editing experience built around how marketers actually work from brief to variant to activation without a specialist at every step.

The brief for Horizon Canvas: build a scalable content operations platform capable of supporting multiple concepts, multiple formats, AI-powered campaign creation and adaptation, image workflows, video workflows, and future generative capabilities without requiring a creative specialist at every step.

The Platform Story

Users arrive after generation. That's where the real work begins.

After a marketer writes a brief and triggers AI generation, they land in Horizon Canvas with multiple concepts. Exploration, selection, adaptation, campaign preparation: everything from this point forward is what Horizon Canvas is designed for.

After generation, users explore concepts in Grid View, select one, and adapt it across channels and formats. Horizon Canvas owns the entire post-generation journey — from concept to campaign-ready asset.

Performance marketers don't arrive at Horizon Canvas with a blank canvas. They arrive with AI-generated content that is almost right. The platform's job is to close the gap between "what AI generated" and "what's ready to activate." Every design decision in Horizon Canvas was made with this context in mind.

Scale + My Role

F500

Enterprise customer tier: GM, Disney+, Unilever, Ulta, Xfinity, Best Buy, Paramount+

AI content adaptation capabilities: Expand, Reframe, Crop, Timeline, Inline lightweight

Unified interaction framework across the entire platform

H1 '26

Design cycle — this project

Horizon Canvas was a multi-designer effort. The areas below reflect where I led, contributed to, or influenced design decisions, not sole ownership of the entire surface.

My Contribution — Overview

Shaped the post-generation content operations workflows that bridge AI generation and campaign activation. Not individual features in isolation, but the platform logic connecting them.
Led GenExpand end to end: the capability that lets a marketer extend an AI-generated image, review alternatives, and commit without a designer.
Established the unified interaction framework, the system that turned disconnected AI capabilities into one coherent platform experience and scales automatically as new capabilities ship.
Partnered on Reframe — translating research insight (marketers think in channels, not dimensions) into a workflow that adapts content for Meta Stories, LinkedIn, and Display without showing a single pixel number
Partnered on Timeline editing, making sure video scene management shared the same interaction language as image operations rather than a separate paradigm.
Led cross-surface consistency: one interaction language across image and video content types, carrying research insights into design decisions throughout.

GenExpand (Led)End-to-end design: workflow, review UX, AI trust model, error states. This was my primary ownership area.

Entry Point System (Led)Proposed the unified hover action bar and drove alignment across feature teams on the shared pattern.

Reframe (Contributed)Worked with other designers on channel-first format adaptation: preview states and the iteration model.

Timeline Editing (Contributed)Contributed interaction patterns for scene management and clip control alongside the team.

Content Canvas Design (Collaborated)Part of the design team that shaped the content canvas: surface structure, panel system, layer interaction.

Cross-Surface Consistency (Led)Led interaction consistency across image and video content types, so the platform speaks one language regardless of which feature team built it.

Success Criteria

Defined before design began. These shaped what got included, what got deferred, and how progress was evaluated.

Marketers can adapt content without a specialist

Format changes, image adaptation, and content operations available without routing back to a creative agency.

AI content operations have explicit human checkpoints

Every generative operation requires user review and selection before committing. AI doesn't apply changes automatically.

One entry point pattern, not many

Users learn one way to discover content operations. New features inherit that pattern, with no additional learning cost.

Video and image content share one environment

Timeline and image operations (Expand, Reframe, Crop) in the same content canvas. No context switch between content types.

Error states are recoverable

When AI generation fails, the user has a clear path back with enough context to understand what happened and what to try next.

Engineering can build from specs alone

Every interaction state (hover, generating, review, selected, error) fully specified across all five content operations.

Product Principles

Exploration and execution require different environments

Comparing concepts requires a different UX than adapting one. Grid View and the editing canvas serve different cognitive modes. Collapsing them into one surface serves neither well.

Discoverability must survive complexity

A surface with video layers, image layers, and multiple AI operations needs one consistent entry point pattern. Hover-based and contextual, not a different toolbar per feature.

AI actions need human checkpoints

Generative expand doesn't replace judgment. The workflow generates alternatives, surfaces them, and leaves selection to the marketer. AI that commits changes without review erodes brand trust.

Consistency is the only thing that scales

A content operations platform with 12 capabilities can't afford 12 interaction paradigms. One system, applied consistently, means every new AI capability reduces friction instead of adding it.

Problem Discovery

GenStudio's research org (ADRS-DX) ran multiple studies in parallel with the design cycle. That research shaped the platform at every stage, from the overall strategy down to how individual capabilities were designed.

Who is Mo

n=8 interviews · Sep 2024

Foundational persona: Modern Marketer archetype. Core pains: creative dependency, tool sprawl, review bottlenecks. AI Enthusiast vs. AI Conscript split.

From Generation to Scale

n=7 interviews + prototype probe · March 2026

Pre/post generation mental models. Explore and scale are distinct cognitive modes. The transition must be explicit and user-controlled.

Paid Media Foundational Work

~10 interviews · May 2025

Performance marketers have two dials: spend and delivery. GenStudio adds a third: create and test content without going back to creative.

Horizon Canvas Template Config

n=5 enterprise marketers · Feb 2026

Channel-first mindset confirmed. Core editing tools must be visible, not buried in overflow menus. Top-down hierarchy (concepts → assets → elements).

Customer Feedback Analysis

n=150 responses · March 2025

Creative control in the creation surface was a top-10 theme. Designers want more layout flexibility and image control, which directly shaped the editing surface scope.

Design vs. Template Study

n=50 quant + n=20 qual · March 2026

As AI blurs the line between designing and templating, terminology for AI-assisted editing needs to stay grounded in user language, not engineering language.

Key Findings

Exploration and scaling are distinct cognitive modes, and the UI must honour that

Users iterate on a concept until they trust it, then shift to scaling it across formats and channels. The two modes require different affordances. One surface trying to serve both ends up serving neither well.

Performance marketers think in channels, not dimensions

Reframe's job isn't "resize to 9:16." It's "adapt this concept for Meta Stories, LinkedIn, and Display." The UI had to work in the marketer's mental model, not the designer's spec vocabulary.

AI content can't go live as-is. Review is the real design problem.

Brand tone, composition, image quality: AI output requires human review before it's marketing-ready. The platform's job is to make that review fast, not remove it.

Core content operations must always be visible

Template config research found that hiding operations in overflow menus increased cognitive load and reduced perceived control. Entry points need to be surfaced, not buried.

Before Horizon Canvas

Every format change required a round-trip to a creative agency
AI-generated assets needed specialist QA before use
No consistent content operations platform: each capability had its own pattern
Video and image content had completely separate workflows
Post-generation workflow had no dedicated design

After Horizon Canvas

Performance marketer adapts format directly via Reframe in the content canvas
GenExpand generates options; marketer reviews, compares, selects
One hover-based entry point pattern across all content operations
Video and image workflows unified in one content canvas
Post-generation workflow is the primary design surface

Ecosystem Map

Horizon Canvas sits at the centre of a complex ecosystem. Users, systems, and data flows all converge on one content adaptation platform that connects AI generation to campaign activation.

Key Product Decisions

Three decisions shaped how Horizon Canvas connects AI generation to campaign activation. Each required evidence-based alignment across product, engineering, and leadership.

Decision 01

Treat concept exploration and content adaptation as two distinct cognitive modes, not one surface

Considered

Unified surface: edit inline from Grid View

Fewer navigation steps
Simpler architecture to build
Industry precedent (Canva-style)

Chosen

Separate surfaces: Grid for exploration, dedicated canvas for execution

Exploration requires comparison; editing requires focus
AI panels need dedicated canvas space
Video timeline can't exist inside a card grid

How it resolved: "Generation to Scale" research confirmed two distinct cognitive modes: comparison and execution. A prototype showing the cognitive load of editing an Expand prompt while surrounded by 12 other concept cards settled the argument. The surface split follows the user's mental model, not engineering convenience.

Decision 02

Anchor the product experience around what happens after AI generates content

Initially Proposed (Product)

Keep the product in the generation view: all adaptation inline within Grid

Fewer navigation steps for the marketer
Consistent with lightweight text + swap operations
Simpler product architecture to ship

Returned to after Jan 2026 Review

Design a post-generation workflow: dedicated space for exploring, adapting, and scaling content

Generation and adaptation are different mental modes. Treating them as one surface serves neither.
AI review states (4 Expand alternatives, channel previews, video timeline) require space the Grid can't provide
Creates a scalable foundation for future AI capabilities

How it resolved: Video was the forcing function. Timeline management and scene control physically cannot exist inside a card grid. Once that was on the table, the same principle applied to every AI operation requiring review states: GenExpand alternatives, Reframe channel previews. The result was a product clarity win as much as an architectural one. The post-generation workflow became the product, not a sidebar to generation.

Decision 03

One shared interaction system over per-feature UI fragmentation

Considered

Per-feature toolbars: each capability designs its own

Each team ships independently
Feature-specific UI can be optimised locally
No cross-team coordination required

Chosen (proposed by Mariam)

Single hover action bar + contextual layer-level bars

One mental model for all content operations
New features inherit discoverability automatically
Consistent visual weight, no competing chrome

How it resolved: Template configuration research (n=5, Feb 2026) showed that inconsistent entry points increased cognitive load and reduced perceived control. I proposed the unified hover bar as a shared component: each feature team adds their actions as slots in one pattern rather than designing a new discovery mechanism. The tradeoff: less custom chrome per feature. The gain: one amortised learning cost across the entire platform.

Feature Deep Dive

GenExpand: From AI Output to Format-Ready Asset

An AI-generated image is rarely campaign-ready as-is. GenExpand closes that gap: a performance marketer extends image boundaries with a natural language prompt, reviews alternatives, and commits to a final result without going back to a creative agency.

My Contribution: GenExpand

Designed the GenExpand workflow end-to-end. My primary ownership area.
Designed the review experience: thumbnail grid, hover preview, comparison, selection before commit
Determined prompt placement: brief-style before generation, not refinement-style after
Designed all error states and recovery paths (generation failure, content policy block, quality threshold)
Worked with engineers on generation state design: loading states and progress feedback

GenExpand prompt entry. Brief-style: write the direction first, then see all generated alternatives.

The Workflow

User hovers image in the content canvas

The Image layer inline bar appears. "Expand" is visible in the action set alongside Reframe and Crop.

User opens Expand panel and enters prompt

Brief-style prompt: directional language ("extend background to show a summer outdoor setting") rather than specific parameters. The system understands creative intent.

AI generates 4 alternatives

Thumbnail grid appears showing the original plus 4 generated options. Generation runs in the background, with a progress indicator visible without blocking the rest of the canvas.

User reviews in the comparison grid

Each thumbnail shows the full expanded image at a glance. Hover reveals full-size preview. A quality signal indicates AI confidence for each option.

User selects and commits

Selecting an option shows the full-size result in the canvas before committing. A secondary prompt allows a refinement iteration on the selected option before applying.

Non-destructive apply

The expanded version becomes a new variant; the original is preserved. The marketer returns to the content canvas with the expanded asset ready for further editing or review.

GenExpand — Review state, 4-option thumbnail grid

GenExpand review state. 4 alternatives with original for reference. Hover reveals full-size preview before selecting.

Design Challenges

The multi-option review problem

After generation, a marketer sees 4 options and needs to compare them and evaluate brand-fit before selecting. Too much information per thumbnail = overwhelming. Too little = uninformed decision. The design: thumbnail grid for quick comparison, full-size preview on hover, no commit until the user has seen the full result.

Brief-style vs. refinement-style prompting

Where does the prompt live: before seeing options (brief-style) or after (refinement-style)? Testing showed users preferred brief-style: commit to a direction, then evaluate all alternatives. This maps to how marketers brief creative teams: direction first, then assessment.

AI confidence and quality signals

A generated expand that subtly breaks brand guidelines is worse than no output. The design includes a quality signal in the thumbnail grid: not a blocking warning, but enough signal that a marketer pauses on a low-confidence option before approving. Trust must be earned, not assumed.

Error recovery as a first-class experience

Generation failure, content policy block, quality threshold: each failure mode has a distinct message and a recovery path. The user should know what happened and what their next step is without leaving the content canvas.

"The most important interaction in GenExpand is not the generation step. It's the review step. Generation is fast and invisible. Review is where trust gets built or broken."

Feature Deep Dive

Reframe: Channel-First Format Adaptation

A performance marketer doesn't think "I need to resize to 1080×1920." They think "I need this concept on Meta Stories, LinkedIn Banner, and Display." Reframe had to work in their mental model: channel-first, not dimension-first. No designer required.

My Contribution: Reframe

Contributed to Reframe workflow design alongside other designers
Advocated for and helped implement channel-first mental model in the UI (not dimension-first)
Designed preview states: side-by-side channel comparison before committing
Contributed to the iteration model: how users refine an adaptation before approving

Reframe — Channel selection and side-by-side preview

Reframe: channel-first selection. Users choose target channels, AI adapts composition, side-by-side preview before approving.

The Workflow

User selects asset in the content canvas and triggers Reframe

From the Image layer action bar: "Reframe" opens the channel selection panel.

User selects target channels, not dimensions

Meta Stories, LinkedIn Banner, Google Display, YouTube Thumbnail. The system maps channels to the correct dimensions. The marketer never sees numbers.

AI adapts composition for each channel

Intelligent recomposition, not a simple crop. The system understands focal points, brand elements, and text placement, then adapts the layout for each format.

Side-by-side preview across channels

The marketer sees all adapted versions simultaneously and can approve per channel or approve all. Any channel that doesn't look right can be iterated with a prompt refinement.

Non-destructive output

Each approved channel adaptation becomes a new variant. The original is preserved. All variants appear in Grid View for the full concept.

Why Channel-First Mattered

Research across paid media professionals showed consistently that marketers think about their work in terms of channels and placements, not pixel dimensions. A "Meta Stories" placement has meaning: it implies vertical format, full-bleed visuals, 15-second view windows, and specific brand safe zones. A "1080×1920" number has none of that context.

Building Reframe around channels rather than dimensions came directly from this research. It reduced cognitive load, cut down on wrong-dimension errors, and made the tool speak the language of its users instead of its engineering spec.

Feature Deep Dive

Timeline: Video Content Without a Video Editor

Video was already the dominant format in enterprise paid media before GenStudio had any video workflow. Timeline management gave performance marketers the ability to manage scenes, sequence clips, and control pacing directly, without handing off to a video editor or leaving the platform.

My Contribution: Timeline Editing

Contributed to Timeline interaction design alongside other designers
Designed specific scene management interaction patterns: drag-to-reorder, trim handles
Timeline's requirements were a key input in validating the dedicated content canvas decision
Helped ensure Timeline used the same entry point pattern (action bar) as image content operations

Timeline — Content Canvas with Timeline panel open

Timeline management inside the content canvas: scene sequencing, clip trimming, and video layer controls unified with image content operations.

The Architecture

Scene-level objects

Each AI-generated or imported video clip is a scene: a discrete temporal object with a duration, position in sequence, and set of editable properties. Marketers think in scenes ("the opening hook", "the product reveal", "the CTA"), not in frames.

Timeline interaction model

Drag to reorder scenes. Trim handles on each clip for duration adjustment. Scene thumbnails for quick visual identification. The timeline extends across the bottom of the content canvas: it needs the full canvas width, which is why it can't exist in a card grid.

AI video actions in the timeline

Generate new scene, extend clip duration, replace scene content: AI actions surface in the timeline context. The same action bar pattern used for image editing applies here, with video-specific actions available on the Video layer.

Why this required a dedicated content canvas

Timeline management physically requires dedicated canvas space. This was the clearest argument for the content canvas: once leadership saw that video editing was impossible inline, the architectural decision resolved quickly. Timeline was the forcing function that validated the approach for all complex operations.

Timeline — Scene reordering and timeline detail

Timeline detail: scene cards in sequence with reorder handles. AI video actions accessible per scene from the same action bar pattern.

Feature Deep Dive

Unified Interaction Framework: The System That Connects Everything

Five AI content adaptation capabilities. Two content types. Multiple feature teams shipping independently. Without a shared interaction framework, marketers would face a different discovery pattern for every capability. The Unified Interaction Framework solved this, ensuring every future capability would inherit the same interaction language automatically.

My Contribution: Entry Point System

Proposed and designed the unified hover action bar. One of my most significant contributions to Horizon Canvas.
Defined the pattern: single action bar at canvas level + contextual inline bars for Video and Image layer types
Drove alignment across feature teams on adopting the shared pattern
Designed the visual and interaction spec for the action bar system
Established this as the standard: new features that ship inherit the pattern automatically

Entry Point System — Hover state canvas level action bar

Entry point system: the hover action bar appears on canvas hover. At rest, the canvas is clean. On hover, all content operations surface in context.

The Problem Before

Before the unified framework was established, each feature team was independently designing how marketers would discover and trigger their capability. Expand had one pattern. Reframe had a different one. Timeline had its own toolbar concept. The product was heading toward a surface where every new AI capability added to the learning cost. That's the opposite of what a platform should do.

The Architecture

↑

Canvas level: primary action bar (hover)

When a user hovers the canvas, the primary action bar appears. It surfaces the highest-priority content operations available for the current asset type. Clean at rest: no chrome competing with the content.

↑

Image layer: contextual inline bar

When a user hovers the Image layer specifically, the Image inline bar appears with Image-specific actions: Expand, Reframe, Crop, Swap. This is a sub-pattern of the canvas-level bar: consistent trigger, contextual contents.

↑

Video layer: contextual inline bar

When a user hovers the Video layer, the Video inline bar appears with Video-specific actions: Timeline, Trim, Replace, AI video operations. Same trigger mechanism, different content set.

↑

Text layer: contextual inline bar

Text layer hover surfaces text editing actions. Lightweight inline: no panel required for simple text and style changes.

Entry Point System — Image layer hover vs Video layer hover

Image layer shows you can change shows group and specific content hover and sleected states and interaction.

Why It Scales

The unified framework's most important property is scalability. When a new AI editing capability ships, it gets added as a slot in the existing pattern. Discoverability is inherited automatically. Marketers don't pay a learning tax for every new capability the product ships.

This is a compounding investment. The first capability that adopts the pattern earns consistent discoverability. So does the fifth. So does the tenth. Every marketer who learns how to find GenExpand already knows how to find Reframe, Timeline, and Crop, and every future capability that follows. That's the difference between a collection of tools and a content operations platform.

Stakeholder Management

The platform decisions required alignment across teams with competing priorities. Three tensions were significant enough that evidence, not just discussion, was needed before they resolved.

Stakeholder	Their Position	Evidence That Shifted It	Resolution
Product	Inline operations for all capabilities: keep the surface shallow, reduce navigation depth. This was the initial direction before the Jan 2026 review.	Video timeline cannot exist inline, physically. Once this was established, the same principle applied to all AI operations requiring review states or panel space.	January 2026 architecture review. Dedicated content canvas established. Lightweight inline operations kept for simple tasks: a principled split.
Engineering	Per-feature entry points: each squad ships its own toolbar independently, without cross-team coordination overhead.	Template config research showed fragmented entry points increased cognitive load. A shared component reduces QA scope and future engineering complexity, not just design complexity.	Unified action bar spec as a shared component. Each team adds feature-specific action slots: less custom work per team, one consistent pattern for users.
Creative Leadership	Hover-triggered patterns feel restrictive, wanted more persistent controls visible for power users who work fast.	Interaction testing showed power users found hover patterns faster once learned: no hunting, actions appear in context where needed.	"Always visible" action bar mode added as a user preference for Phase 2, acknowledging the use case without compromising the default clean state.

Validation

Decisions were tested through a combination of formal research and structured internal review, not assumed correct after the first prototype.

Prototype probes: 1:1 with enterprise marketers (n=7, Generation to Scale study) Template configuration research (n=5, Feb 2026) Stakeholder walkthroughs: product, engineering, creative leadership January 2026 architecture review: leadership sign-off Cross-functional design critique sessions Engineering feasibility reviews: interaction state by state

The January 2026 leadership review was the critical moment. Leadership approval doesn't equal user validation, but the architectural question (inline vs. a dedicated content canvas for post-generation workflows) was a product strategy decision that needed explicit organisational alignment before design could move forward. User research informed the proposal. Leadership alignment made it buildable.

Reflection

The most challenging part of Horizon Canvas wasn't designing any individual feature. It was designing for a team. The entry point system worked precisely because it was a shared agreement among designers about how the product would behave, not just a solution for users. Getting that agreement required understanding what other designers were building and why, and making the case for a shared pattern before anyone's feature was locked.

The January 2026 review was a lesson in how architectural decisions get made in practice. A proposal I'd contributed to was returned to the table. The instinct is to defend it. The more productive move was to understand what the concern was actually about: video was the real answer. Once that was on the table, the right direction became obvious to everyone. The best outcome wasn't winning the argument. It was the team landing on a better, more defensible architecture together.

Designing GenExpand taught me that the hardest part of AI product design is the review state. Generation is fast, impressive, and invisible. Review is where real decisions happen: comparison, evaluation, the moment trust gets built or lost. If I got the review state right, the feature felt trustworthy. If I got it wrong, no amount of impressive generation would make up for a bad selection experience.

What I Learned

Shared interaction patterns require alignment before they can be designed. The entry point system was an alignment problem before it was a design problem. Understanding what each team was building, and making the case for a shared standard before features were locked, was harder than designing the pattern itself.

In AI product design, the review state is the product. Generation is infrastructure. What earns user trust (or destroys it) is what happens between AI output and user commitment. Every AI operation I worked on had more design complexity in the review state than in the generation trigger.

Platform decisions compound. The choices that determine whether a product becomes a coherent ecosystem or a set of disconnected tools are made early: before features are locked, before engineering has started. Getting the foundation right is worth more than shipping faster on an inconsistent base. One consistent interaction framework means every new AI capability the product ships makes the experience better, not more complicated.

Figma Adobe Spectrum Prototyping User interviews Prototype probe testing Cross-functional alignment Interaction specification AI workflow design