How I AI: Gamma's 3-Step AI Workflow for Global Feedback, Art Direction, and Hiring

A thirty person team is using AI to review hundreds of multilingual support responses, generate production-ready illustrations, and standardize hiring materials without adding layers of process. The interesting part is not the tooling itself. It is how Gamma keeps human judgment in the loop while moving much faster.

In this episode of How I AI, Zach analyzes 550 responses in more than a dozen languages, develops an illustration through rapid Midjourney iterations, removes its background with Replicate, and uses a Claude Project to draft a fictional job description.

Gamma Head of Design Zach Leach walks through the workflows behind those outputs. At the time of recording, roughly 60% of Gamma's users were international and did not speak English. AI widened the team's reach, while interpretation, art direction, and hiring decisions remained human.

Turn multilingual feedback into product signals

Gamma collected 550 responses about its AI image editor in more than a dozen languages over the course of a week. The responses included complaints about distorted people, failed edits, and model quality, mixed with praise for features that worked surprisingly well. Zach wanted a view across the entire dataset instead of relying on a quick skim of English-language feedback.

Before this workflow, Zach would have skimmed a few dozen responses, mostly in English. The multilingual feedback-analysis workflow shows how deep research widens the sample while keeping source review and qualified language judgment in the loop.

Prepare the research dataset

Zach exported the responses into a single file and uploaded it to ChatGPT using deep research mode. He specifically wanted reasoning and classification, not lightweight keyword grouping. Teams using this approach still need to handle the operational side carefully: remove unnecessary personal data, verify workspace retention policies, and limit access to people directly involved in the research.

Ask for a product-team level analysis

Instead of requesting a generic summary, Zach framed the prompt around product decisions. He asked ChatGPT to identify what users loved, what consistently failed, why certain edits succeeded, and where the product experience broke down. That framing mattered because the goal was not sentiment reporting. It was roadmap input.

"This is some feedback we've received about our AI Image editing feature. I want you to analyze the feedback and find where we are doing poorly, and where we are doing well. Break down for our product team, what kinds of things we are doing well and why, and what kinds of things we are doing poorly and why. What do people love? What do people hate? Where can we improve?"

The deep research run took about 19 minutes. ChatGPT translated responses, grouped recurring complaints, and generated themes with supporting examples. Zach noted that his first attempt without deep research produced something much shallower: a basic Python-style keyword analysis that missed nuance and treated the feedback more like spreadsheet cleanup than product interpretation.

The improved workflow surfaced more useful distinctions. Users liked image upscaling. They struggled with multi-step edit requests. Complaints about extra limbs and distorted people appeared across languages. Zach could also inspect translated excerpts directly instead of relying only on aggregate summaries.

That said, translation and clustering still flatten nuance. The workflow works best as a discovery layer, not final truth. Teams still need to review source responses, sample multiple languages with qualified reviewers, and treat AI-generated counts and categories as directional rather than definitive.

Zach first received a shallow keyword analysis. Specifying deep research and giving the model a clearer analytical brief produced a more useful report.

Convert the analysis into a working presentation

Once the research was complete, Zach copied the findings into Gamma and asked it to generate a presentation with charts and graphs. The resulting deck became a shared discussion surface for product and engineering rather than a static research document.

The analysis highlighted both feature wins and UX problems. Users responded positively to upscaling quality, while multi-step image edits often failed because the model handled only part of the request. That led Zach to think about UX interventions rather than model swaps alone. For example, the interface could detect complex prompts and suggest splitting them into smaller steps automatically.

The workflow compressed what would normally be days of translation, synthesis, and slide preparation into a much shorter cycle. But the final interpretation still required product judgment. The team still had to validate chart inputs, review translated excerpts against source feedback, and decide which findings actually justified roadmap changes.

What stands out in this workflow is not just the summarization. It is the ability for a small team to stay close to a global user base without creating a separate research operation. Zach described AI as a way to collect more freeform feedback, from more users, in more languages, because the cost of sorting through the responses has dropped dramatically.

Use Midjourney as an art direction system

Gamma’s rebrand leans into surreal, airy, colorful imagery with floating animals, vivid gradients, and playful compositions. Zach and Gamma’s creative director built reusable Midjourney references so the team could generate assets that felt consistent with the brand instead of starting from scratch every time.

The important distinction is that they are not treating generated images as final artwork by default. The AI handles exploration and variation. Humans still decide what fits the product and what actually communicates the intended feeling.

Build reusable brand references

The team relies on Midjourney’s Style Reference parameter and a shared set of prompt patterns to keep outputs aligned with the brand system. Zach described it almost like a loose design kit for image generation. Instead of relying entirely on natural language prompts, the references narrow the visual range toward Gamma’s established look and feel.

That consistency matters more when multiple people across the company are generating assets. The references help outputs feel related, but originality, composition quality, rights review, and brand fit still require human review.

Follow the ideas that unexpectedly work

For an empty state inside Gamma’s image editor, Zach started exploring visual metaphors for transformation. Rather than locking into one concept, he moved quickly through many directions and kept refining whichever ideas felt promising.

1. He started with surreal painting concepts but decided they did not communicate the interaction clearly enough.
2. He experimented with people chatting to represent conversational image editing, then moved on again when the compositions felt generic.
3. He explored an apple split between green and red to represent transformation and editing states.
4. One generated variation unexpectedly introduced a bird, which became the breakthrough direction for the final concept.

Zach iterated on the bird's color, composition, and split-image structure until it matched Gamma's visual language. I compared the speed with compressed agency review cycles. The on-brand illustration workflow captures the path from style references and selection through background removal and production review.

Prepare the image for production use

Once Zach selected the illustration, he still needed a transparent version for use inside the interface in Figma. The generated image included a full background, which was visually useful during exploration but impractical for production UI work.

He uploaded the image to a background-removal model on Replicate and exported a transparent PNG. Zach described the workflow as fast and reliable enough that it replaced older manual techniques like tracing masks by hand.

The final image slightly overflowed its card boundary inside the product, reinforcing Gamma’s broader visual style of breaking beyond traditional slide layouts. But even at this stage, generated assets still required review. Teams need to inspect edge quality, confirm image resolution, verify source and model rights, document provenance where appropriate, and test how the asset behaves inside the actual interface.

Create consistent job-description drafts with Claude

Gamma also uses AI to standardize hiring materials. Instead of asking every hiring manager to write job descriptions from scratch, the company created a Claude Project that generates first drafts based on approved examples.

The goal is consistency, not automation of hiring decisions. The system helps standardize tone, structure, and formatting, while humans still define the role, expectations, compensation, and evaluation criteria.

Load approved examples into Claude Projects

The team uploaded several strong existing job descriptions along with instructions describing Gamma’s preferred voice and structure. Zach showed how the project uses those examples almost like style conditioning for future drafts.

For real company use, those examples should be current, legally reviewed, and stripped of sensitive employee or candidate information. The quality of the outputs depends heavily on the quality of the source material.

Generate a draft role in the company voice

To demonstrate the workflow, Zach intentionally tested it with a ridiculous fictional role:

make a job for head of popcorn

Claude generated the fictional role in Gamma's existing voice and structure. The job-description workflow shows how approved examples create a consistent first draft while hiring, legal, compensation, accessibility, and people teams retain review.

Review the draft and publish it in Gamma

After generating a draft, a hiring manager can edit the content, move approved copy into a Gamma template, and publish a careers page quickly. Zach emphasized that the workflow saves time on repetitive writing while helping the company maintain a consistent hiring voice.

Human review still matters at every stage. Before publishing a real role, hiring, legal, accessibility, compensation, and people teams should review qualifications, inclusive language, location requirements, claims, and application instructions.

AI expands the team’s range, not its authority

Across all three workflows, the pattern is consistent. AI handles scale, variation, translation, formatting, and first drafts. Humans decide what deserves trust, what fits the brand, what becomes product direction, and what actually ships.

That division of labor is especially noticeable in the design workflow. Midjourney generated hundreds of possibilities, but Zach still relied on taste and critique to identify the one image that actually communicated the product experience. The same thing happened in research. ChatGPT surfaced patterns, but humans still interpreted whether those patterns justified UX or roadmap changes.

The strongest parts of these workflows are the bounded inputs and the fast iteration loops. Gamma feeds the models clear source material, asks for inspectable outputs, and keeps a specific human owner responsible for the final decision.

That approach is worth copying for teams handling large-scale customer feedback, brand systems, or repetitive internal documentation. The AI meaningfully reduces tedious work and expands what a small team can cover. What it does not remove is judgment. Translation still loses nuance. Generated visuals still need taste and review. Hiring copy still needs legal and organizational scrutiny. The leverage comes from compressing the expensive middle of the process, not from eliminating the humans at the end.