Research
What Deserves Confidence
Three research systems that helped a team move faster, and know when an idea had actually earned its confidence.
Haim Repael · 4 min read

STAR AT A GLANCE
- Situation
- Product teams rarely have time to resolve the tension between conviction and curiosity, and AI makes research-shaped nonsense easy to produce.
- Task
- Get to the important questions earlier, without letting speed pass for knowledge.
- Action
- A synthetic panel grounded in real customer evidence, a closed-garden research tree, and a Voice of the Customer graph.
- Result
- 87% alignment between the synthetic panel and the human study, a first plugin version built in under three weeks, and partner critique that moved from "you misunderstood us" to "this could be better."
I've always liked knowing why. Not collecting facts for their own sake, but the moment when something vague starts acquiring edges. One answer exposes a better question, which sends you somewhere you hadn't planned to go. I can lose hours that way.
Design makes that instinct complicated. You can research forever and still be wrong. You can also move fast, trust your experience, ship something beautiful, and discover in fifteen seconds that someone uses it in a way nobody in the room predicted.
That moment never made me like research less. It made me like it more. The user's reaction is the receipt.
SituationConviction versus curiosity, on a deadline
Excellence lives between two impulses: enough conviction to make something, and enough curiosity to find out whether that conviction was earned. Product teams rarely give you unlimited time to resolve the tension.
And AI added a new risk. It made research ridiculously fast, and it made it dangerously easy to produce research-shaped nonsense: ten beautifully written reports quietly citing one another, an LLM's confidence passing for evidence.
TaskResearch faster, without shortcuts
Not shortcuts around research. Systems for reaching the important questions earlier, while keeping one rule intact: a claim only gets promoted if it can be traced back to evidence I trust.
ActionThree systems
1. One hundred users who didn't exist
We were redesigning a personalized dashboard when our ideal customer profile changed. From a design seat, that meant some of what we thought we knew about our users was no longer good enough.
Ofri, the product designer working with me, and I went back to the personas. We collected everything we could reasonably know: behavioral data with our analysts, previous research, customer conversations from CX, and more calls transcribed with a local tool I built. We ended with five primary personas, rich enough in goals, workflows, fears and decision criteria that a persona could disagree with us.
Then I had an idea: what if these personas could do something? Not a chatbot pretending to be "Sarah, 38, marketing manager." A panel. I built it: one hundred synthetic users grounded in those five personas, distributed according to the real composition of our customers. We gave them the dashboard, structured questions, open questions and tasks, and looked for disagreement, repeated friction and missing information.
The panel surfaced a missing user story, a workflow that made sense for one persona and almost none of the others, an edge case we had wrongly treated as marginal, and, most valuable of all, work that looked interesting but did not deserve to be built.
But I never trusted it enough to stop there. We put the same work in front of real customers.
2. A research tree, inside a closed garden
When I started on an accessibility plugin for Figma, I didn't lack users. I lacked a mental model. Every answer opened three more questions. Then five. Then twelve.
So research became a tree: one large question, broken into branches, sent through different tools and source sets, contradictions kept visible instead of averaged away, then collapsed back into one synthesis I could use.
The sources only got in because I let them in. Primary documentation beat summaries. Claims had to be traceable. Only later did I recognize the established methods inside what I'd built: triangulation, thematic analysis, affinity mapping. I ran into the problems they were invented to solve, then went back to the literature and stole the parts that made mine better.
3. The customer already told us
Companies have an absurd amount of customer knowledge, rarely in one place. Support has one piece, analytics another, sales hears the objection first. Then someone walks into a meeting and says "customers want this." Which customers? How many? Did they describe a problem, or did we supply the solution?
I built a Voice of the Customer repository that connects each piece of feedback to a segment, a journey stage, a product area, observed behavior, frequency and urgency. Not a quote library, which lets whoever has the best quote win the meeting. A graph of relationships, prioritized by prevalence, severity, confidence in the evidence and effort, never by popularity alone.
ResultUseful, not human
- 87% alignment between the synthetic panel and the human study, across the measures we compared. The agents had not become users. They had become useful. We copied the model to other product areas: synthetic first while being wrong is cheap, humans when the question deserves humans, production because nothing else reproduces a real Tuesday afternoon.
- One insight changed what I built: design-time accessibility is not DOM accessibility. In Figma, a rectangle might be a button or decoration. The problem was classification before rule checking. The first working plugin version took me just under three weeks.
- Hundreds of pieces of feedback on the Services experience became a handful of underlying needs and one coherent flow. When partners saw it, the response was overwhelmingly positive, and critique moved downstream, from "you misunderstood us" to "this solution could be better." I'll take that trade every time.
What travels
I still trust instinct. Some decisions are reversible, cheap, or simply taste. I just don't give instinct diplomatic immunity.
AI increases the amount of research I can touch. It does not remove my responsibility for deciding what deserves belief.
A hundred synthetic users can be wrong together. Twenty reports can come from one weak source. Eventually you stop researching and ship. Someone you have never met opens what you made. They don't know about the panel, the tree or the graph. They shouldn't. They just use it, and I watch what happens.