How I AIGamma’s head of design on using AI to synthesize feedback and generate on-brand imagery | Zach Leach
CHAPTERS
- 0:00 – 3:34
Gamma’s global user base and the challenge of multilingual feedback
Claire and Zach set the stage: Gamma is a highly international product, which creates a real challenge for design teams trying to learn from customer feedback at scale. They frame the episode around using AI to stay close to users despite language and volume barriers.
- •Gamma’s users are majority international, many in non‑English languages
- •Design needs to synthesize diverse input into actionable product improvements
- •AI is positioned as a way to scale research and brand execution
- 3:34 – 5:43
Feature walkthrough: Gamma’s AI image editor and how feedback is captured
Zach demos the AI image editing feature inside Gamma, showing how users chat to request edits and then submit feedback when results miss the mark. This feedback stream becomes the raw material for quality evaluation and model/UX decisions.
- •AI image editing via chat-based instructions (e.g., “add caramel drizzle”)
- •Users can flag bad outputs (e.g., “too much drizzle”)
- •Gamma collects free-form feedback tied to edit performance
- •Typical generation issues appear (extra arms/fingers, odd artifacts)
- 5:43 – 7:25
Uploading 550 responses into ChatGPT Deep Research to synthesize insights
Zach explains how he uses ChatGPT’s Deep Research by uploading the feedback file and prompting for structured analysis. The tool takes ~20 minutes but returns translations, themes, and a breakdown of what’s working vs. not working.
- •~550 individual responses collected in a week
- •Deep Research translates and summarizes across many languages
- •Outputs include “what people love,” “what people dislike,” and trends
- •Deep Research asks clarifying questions about what analyses to produce
- 7:25 – 9:32
Turning qualitative feedback into structured data and a shareable deck
After Deep Research produces an initial synthesis, Zach describes how he can push further: classify each row, generate a structured spreadsheet, then visualize patterns. He then pastes the findings into Gamma to auto-generate a presentation for internal alignment.
- •Follow-up queries: row-level classification, categories, ratings
- •Structured data can be exported for charts/graphs and further analysis
- •Gamma converts pasted research into a presentation draft
- •Deck can include quotes/citations and UX improvement ideas
- 9:32 – 11:41
Before AI: manual sampling and basic scripts vs. Deep Research depth
Claire asks how Zach would have handled this previously, and he admits he would have sampled a small subset—mostly English. Zach contrasts Deep Research with earlier attempts using Python keyword matching or per-row prompting, emphasizing the insight quality difference.
- •Without AI, he would have reviewed ~20 responses manually
- •Prior approach: Python keyword matching or simple automation
- •Deep Research provides richer classification and reasoning
- •Tradeoff: Deep Research takes time but yields higher-quality synthesis
- 11:41 – 14:24
What the analysis revealed: paid vs free quality gap and UX roadmap ideas
They discuss how the feedback can be sliced in different ways and Zach shares a concrete segmentation: paid (Pro) vs free model quality. He also describes actionable UX/product insights like better handling of multi-step edit requests.
- •Segmentation example: Pro vs free users showed ~5% rating difference
- •Better models correlate with better outcomes
- •Key complaint: multi-step edits often partially fail
- •UX idea: detect multi-step prompts and guide users to split/clarify steps
- 14:24 – 16:33
Why this matters for design teams: scaling research in an underfunded area
Claire reflects on how user research is often under-resourced, and how AI changes the economics of listening to customers. Zach highlights how teams can collect more free-form input now and “sort it out later” with AI-assisted synthesis.
- •Research capacity is typically constrained on design teams
- •AI enables large-scale analysis that used to be tedious/expensive
- •Teams can gather more unstructured feedback with less fear of overload
- •Aggregate synthesis helps teams stay closer to the customer
- 16:33 – 19:00
Scaling brand and art direction with Midjourney inside a small team
Claire transitions to Gamma’s rebrand and asks how a ~30-person company maintains high craft and consistent art direction. Zach explains how Midjourney helps them generate on-brand assets quickly, akin to having an always-on art department.
- •Rebrand goals: imaginative, airy, surreal, fun
- •Traditional art production is slow and requires ramping artists on style
- •Midjourney becomes part of a repeatable workflow for consistent assets
- •Small team can ship high-craft visuals faster
- 19:00 – 23:51
Midjourney prompt evolution demo: finding the right metaphor for “transformation”
Zach walks through iterative generations for an empty-state illustration for the image editor. He shows how the concept evolves from paintings to chat motifs to a split-form “transformation” idea—ultimately landing on the half-and-half bird visual.
- •Use case: an “empty state” that teaches the image editor feature
- •Iteration path: painting → chat concept → transformation metaphor
- •Serendipity in generation sparks new directions (animal/bird emerges)
- •Prompting becomes progressively more specific to converge on the final image
- 23:51 – 25:41
Production-ready asset workflow: removing backgrounds with Replicate and placing in Figma
After selecting the winning Midjourney image, Zach demonstrates how he quickly removes the background using a purpose-built model on Replicate. He then drops the transparent asset into Figma to create a polished, on-brand UI composition.
- •Problem: Midjourney output includes unwanted background
- •Tool: Replicate model for high-quality background removal
- •Fast pipeline to generate a transparent cutout for design comps
- •Final UI integrates the illustration with Gamma’s “break out of the slide” vibe
- 25:41 – 29:26
Style references (SREF) and an internal “kit” for brand-consistent generations
Claire probes the mechanics behind consistent outputs, and Zach explains Midjourney style references and personalization. They describe creating a lightweight internal kit—keywords + SREF + shared patterns—so many teammates can generate assets that match the brand.
- •SREF + personalization helps lock outputs to a recognizable brand style
- •Brand system includes keywords and shared prompting patterns
- •The “kit” is socialized internally so others can generate on-brand imagery
- •Without the right constraints, early generations drift off-brand
- 29:26 – 32:47
AI for hiring operations: a reusable Claude Project for consistent job descriptions
Zach shares a separate workflow: using a Claude Project preloaded with example postings and instructions to draft new roles in Gamma’s voice. The output is then formatted and polished using Gamma templates, saving time while keeping quality consistent.
- •Claude Project contains examples + guidance for tone/structure
- •Any hiring manager can generate a role description from inputs
- •AI gets the draft ~80% there; humans still edit and finalize
- •Gamma templates are used to format postings into a polished layout
- 32:47 – 36:20
What humans keep: craft, fun, and personality (plus a lighthearted wrap-up)
Claire closes by asking what Zach wants to “cling onto” as AI expands—he answers: making products fun and engaging. They end with humorous anecdotes about prompting etiquette and Zach’s personal Deep Research use, then wrap the episode.
- •Zach’s core human value: injecting fun, tone, and engagement into UX
- •Prompting style: playful “gentle parenting” vs. being harsh
- •Personal use case: Deep Research for niche topics (pope conclave), imperfect prediction
- •Final thank-yous, subscribe/review call-to-action
