Every so often someone describes their job in a way that reorganizes how I think about mine. Rie McGwier, Staff Researcher & Insights System Architect at Figma, did that at our August Industry Spotlight.
I spend most of my time at Sprig thinking about how research infrastructure should work: what to automate, what to template, where the guardrails go, and who gets the keys. Rie's talk was a reminder that none of that comes first. Infrastructure is always downstream of one person running the loop by hand, unglamorously, long enough to learn what standards should actually look like.
Rie presented four lessons they learned from building a continuous sentiment measurement system from the ground up, and below is what especially resonated with me from each.
Allies get you moving while accelerants get you funded. Don't confuse the two.
Before a program has proven return, two things move it forward: allies who buy the vision early, and accelerants that close the gap between vision and reality. Rie emphasized keeping these strictly separate, and warns against mistaking the enthusiastic people for the powerful ones.
Allies believe in the direction before a metric exists to justify it. Rie found theirs by hunting a specific profile: teams with new products launching, low survey response rates, and audiences that are hard to reach qualitatively. Those teams feel the pain most, so they're the most willing to try something unproven. But an ally is not a sponsor; allies give you morale and momentum while sponsors give you political capital. If you treat an enthusiastic ally like a sponsor early on, you'll over-promise against backing you don't have.
You can't automate a process you've never personally felt
One of my key takeaways was this next rule: do the work by hand until you understand every dimension of it, and let repetition tell you when to stop. Three times through the same motion is the signal to automate the process.
The loop Rie ran by hand was tedious: launch a survey, download the CSV, format it in Sheets, paste the visuals into FigJam, sit with the teams, and repeat. But doing it manually surfaced which pieces genuinely needed to be reusable and where human error was most likely to creep in. These observations became the build plan for what to systematize, which could only be produced by being in the system. What Rie's stack looks like now has built up progressively from that understanding. Their stack now includes Claude Code for self-serve exploration, dbt models with a semantic layer as the single source of truth, reusable Hex components, datasets enriched with behavioral and firmographic columns, and hourly automated ingestion from Sprig.
“A large language model does not know what good looks like. You do, because you were the system first.”
In this model, Rie is on the loop, not in it, and they earned that position by being in it first. This is why the Sprig feature I care about most isn’t the speed; it's that when the Synthesize Agent hands you a theme, you can still trace it back to the verbatim that produced it. Fast research only stays good research when the person on the loop can check the work.
Think in layers, ship in slices
Rie thinks about their program in four layers: the signal, the meaning, the experience, and the extension. They argue you should never deliver all four at once.
A natural temptation is to reveal the whole stack to prove how complete it is, but shipping in slices instead allows stakeholders to react before you spend a quarter investing in a single layer.
Rie started their system with survey data, where they had the most end-to-end control, and only later added harder-to-govern sources like Zendesk, social, and agent feedback. Team-specific Hex dashboards came next, then the Claude skill and the Slack bot, once their value was obvious enough that nobody needed convincing. Each slice earned the justification to invest in the next one.
They shared three questions that keep every project honest: What visible improvement will people actually feel? What foundational investment sits underneath it? And what's the minimum lovable experience, not just the minimum viable one? A slice has to clear all three before it's worth building. That third question is the one most research programs skip, and it's why so many internal tools sit unused.
Be like water: find the crack that exists right now
Rie borrows the image from Bruce Lee of being formless like water. In practice, this means not waiting for perfect conditions to exist and instead finding the opening that appears today and moving through it.
Config made this concrete for Rie. Running the daily pipeline hourly for new product launches would have pushed load times past fifteen minutes so nobody would use it, so rather than waiting for a real fix, Rie used Claude Code to build a producer notebook. The producer notebook runs hourly behind the scenes, pushes to a secure content store, and feeds a lightweight interface for end users. It worked fast, so VP and director-level Figmates were coming back to it daily in the first week, and research ICs started folding it into their own studies.
The conditions were the deliverable
Rie didn't scale by producing more reports; they scaled by making it faster and safer for other people to find the answers themselves. If you're leading an insights program under similar constraints, the first move is smaller than it looks. Pick one team that's flying blind, run one survey, and bring back one verbatim that makes someone lean forward in their chair. That's the spark that buys you the room to build everything after it.
.png)
.png)