How I AIGamma’s head of design on using AI to synthesize feedback and generate on-brand imagery | Zach Leach
CHAPTERS
Gamma’s global user base and the design challenge of multilingual feedback
Claire and Zach set the context: Gamma serves a highly international audience, which means product feedback arrives in many languages and formats. They frame the core challenge for a small design team—staying close to customers while processing feedback at scale.
- •~60% of Gamma users are outside the US, many in non-English languages
- •Internationalization/localization is a major product focus
- •Design decisions must incorporate diverse, multilingual customer input
- •Goal: extract actionable insights without a large research team
Shipping AI image editing in Gamma—and instrumenting feedback loops
Zach demos Gamma’s AI image editing feature and shows how users can request edits via chat (e.g., adding caramel drizzle). Crucially, Gamma captures user feedback on results to learn which prompts and edit types succeed or fail.
- •AI image editing lets users refine generated images inside a deck
- •Users can submit feedback when results are off (e.g., “too much drizzle”)
- •Feedback collection is used to diagnose prompt/model performance
- •Common failure modes include visual artifacts (extra arms/fingers)
The scale problem: 550 responses in a week across many languages
Zach reveals the volume of incoming feedback—hundreds of responses in a single week—making manual review and translation unrealistic. The multilingual nature of the dataset is highlighted as a key barrier to traditional analysis.
- •~550 pieces of feedback collected over one week
- •A substantial share is non-English (multiple languages in top rows)
- •Manual translation and categorization doesn’t scale for a designer
- •Need for a workflow that handles language + classification together
Using ChatGPT Deep Research to synthesize and translate feedback
Zach walks through uploading the feedback file into ChatGPT Deep Research to produce a structured analysis of what’s working and what isn’t. He emphasizes the value of translations, trend extraction, and the ability to ask follow-up questions after the deep run completes.
- •Deep Research run takes ~19–20 minutes for full pass-through
- •Outputs include translated examples plus themes (praise/complaints/trends)
- •Deep Research identifies which prompts/edit types work vs. fail
- •Post-analysis, Zach can ask for row-by-row classification and exportable tables
From analysis to stakeholder-ready output: generating a deck in Gamma
After Deep Research produces findings, Zach copies the results into Gamma to quickly create a shareable presentation for product/engineering. He requests charts and data visualizations to make trends legible and discussion-ready.
- •Paste Deep Research output into Gamma to generate a presentation draft
- •Prompting Gamma to use charts/graphs/data vis instead of photos
- •Creates a fast starting point for product team alignment
- •Pairs qualitative quotes/citations with high-level quantitative framing
Before Deep Research: sampling bias and basic scripts vs. deeper insight
Zach contrasts the new approach with the old: manually scanning a small English-only subset or writing simple keyword-based scripts. Deep Research offers richer categorization and understanding compared to brittle keyword matching or per-row single prompts.
- •Previously, Zach might only hand-review ~20 comments (mostly English)
- •Earlier approach: Python scripts with keyword matching or basic prompting
- •Keyword methods miss nuance and deeper categorization across languages
- •Deep Research reduces pain while improving insight quality
What the analysis surfaced: model tiers, multi-step edit failures, and UX ideas
They discuss concrete insights uncovered by the synthesized feedback, including differences between paid and free model experiences and a major usability issue with multi-step edits. Zach describes potential UX/product responses, like prompting users to split complex requests.
- •Paid vs. free users showed ~5% rating difference (better models → better outcomes)
- •Users frequently complained about multi-step edits failing partially
- •Opportunity: UX that detects multi-intent prompts and suggests splitting steps
- •Also surfaced areas to elevate (e.g., upscaling performing strongly)
Scaling brand on a small team: defining Gamma’s new art direction
Claire transitions to Gamma’s rebrand and the cost of maintaining high-quality brand visuals with limited headcount. Zach articulates the desired style—airy, surreal, vivid—and frames AI as a way to approximate an art department’s throughput.
- •Gamma’s brand aims for imaginative, light, surreal, and fun visuals
- •Traditional approach would require multiple artists and long turnaround
- •Challenge: maintaining consistency across many assets and contributors
- •AI enables faster creation while staying within an art-directed system
Midjourney workflow demo: iterating toward an on-brand empty-state illustration
Zach shows how he iterated through many Midjourney generations to create an empty-state image for the image editor panel. He explains how exploration (apples → birds → split transformation concept) led to the final, on-brand visual that communicates “transformation.”
- •Use case: an educational/empty-state visual inside the image editor UI
- •Iterative exploration: painting → chat motif → transformation concept
- •Serendipity: an “apple” prompt produced a bird, sparking a better direction
- •Prompt refinement becomes increasingly specific to converge on the final image
From ‘agency revisions’ to creative rabbit holes: why rapid iteration matters
Claire highlights how AI changes the dynamics of art direction by making iteration cheap and immediate, avoiding the slow back-and-forth typical with agencies. Zach describes the freedom to chase creative “rabbit holes” until something clicks.
- •AI iteration avoids slow, costly agency-style revision loops
- •Designers can explore many options quickly and decide once they “see the one”
- •Rapid generation supports higher craft and more experimentation
- •Outcome: a stronger, more thoughtful UI detail than a generic empty state
Productionizing the asset: Replicate background removal and Figma integration
Zach explains how he removes the Midjourney background using a specific Replicate model, producing a clean transparent asset for Figma. This replaces older manual approaches like pen-tool masking and accelerates shipping polished UI visuals.
- •Replicate hosts purpose-built models; Zach uses one for background removal
- •Fast, high-quality transparency output suitable for design tools
- •Workflow: generate in Midjourney → remove background in Replicate → paste into Figma
- •Enables polished compositions (e.g., image popping out of a card) quickly
Maintaining brand consistency: Midjourney Style References (SREF) and personalization
They break down how Gamma keeps outputs consistent by using Midjourney style references and personalization developed with their brand agency/creative director. Zach describes it as a loosely shared “kit” that helps the whole company generate aligned visuals.
- •SREF (Style Reference) and personalization steer generations toward brand consistency
- •Brand kit combines reference codes + prompt patterns + shared guidance
- •Without the style system, early generations can drift off-brand
- •Socializing the kit enables non-designers to create usable, consistent assets
A reusable AI workflow for consistent job descriptions (Claude Projects)
Zach shows a separate internal workflow: using a Claude Project seeded with example job posts and instructions to generate new role descriptions in Gamma’s voice. The aim is speed plus consistency so any hiring manager can draft high-quality postings.
- •Claude Project contains examples + instructions to match Gamma tone and structure
- •Generates a job description draft quickly (demoed with a humorous “popcorn” role)
- •Not copy-paste final, but gets teams ~80% of the way there
- •Process can be reused across hiring managers and then formatted into templates
Human craft in an AI-assisted design world: making products ‘fun’ + closing reflections
In the wrap-up, Zach argues the durable human contribution is taste and “finding the fun”—personality, engagement, and delight in product experiences. They end with light discussion of prompting temperament, a personal Deep Research anecdote, and final show outro.
- •Zach’s north star: humans add fun, personality, and engagement beyond automation
- •Prompting style: gentle humor vs. getting “mean” when models misbehave
- •Personal AI use: Deep Research on conclave/papal dynamics (and losing a bet)
- •Conclusion: AI expands capacity, but designers still shape experience quality
