Case study · gitbent
Validating a community in 30 days.
How do you know a community is real before you spend a year building it? You run a 30-day experiment, and you instrument it well enough to tell demand from wishful thinking.
| Timeline | 30-day validation sprint · June–July 2026, within a 90-day experiment. |
|---|---|
| Industry | AI tooling · online community · product. |
| Project solution | A free community and weekly ritual for non-expert AI builders to show unfinished work and get honest, calibrated feedback. |
| Team | Two co-founders, operating as community & validation lead alongside a brand/vision co-founder. |
| Stakeholders | Co-founder, a partner community host (Product Coffee), and the members themselves. |
| Methodology | Continuous discovery (Teresa Torres), lean validation, tiered cohort seeding, funnel + participation instrumentation. |
The challenge
A new wave of people are building real things with AI who never trained as engineers: product managers, marketers, ops leads, curious hobbyists. They have tools that finally let them build, but nowhere to bring the half-finished, slightly-broken work and ask what’s working, what’s not, what would you try next. The polished demo circuit doesn’t serve them; generic Slack groups don’t either.
The open question wasn’t whether we could build a community. Anyone can open a Slack. It was whether the room genuinely existed: real demand, real people, real willingness to show up and engage. And it had to be answered under hard constraints: unpaid, part-time, alongside an active job search, with a co-founder equally stretched. Over-investing before validating was the exact failure mode to avoid.
So the work became a designed experiment: test surface area and demand in 30 days, cheaply, and build enough measurement that the decision to continue, or stop, would rest on evidence, not attachment.
The action
Standing up the community and its front door
I stood up the community and its front door: a tiered seeding motion (warm, adjacent, and cold outreach with distinct scripts), a two-channel comms engine on LinkedIn and Slack, and a weekly anchor ritual, Force Push Friday, where builders show real work against three consistent prompts. The ritual gave the community a heartbeat and a reason to return.
Instrumenting the part most launches skip
Underneath the visible community, I built the part most launches skip: the instrumentation. I tracked the full acquisition funnel, ran a continuous cadence of customer interviews, synthesized them into structured discovery snapshots, and produced weekly readouts against explicit checkpoints, so at any moment we knew not just how many people had joined, but who was actually participating, and why the ones who weren’t had gone quiet.
Discovery over assumption
The instrumentation earned its keep. Seven customer interviews, the cheapest input in the whole experiment, produced nearly every strategic insight we had: a beginner on-ramp problem that was quietly driving early churn, an emerging advisor archetype, and a distinction between two builder types that reframed who the product was even for.
The participation data told an even more important story than the growth data. Acquisition worked: warm outreach and LinkedIn drove a strong top of funnel. But activation lagged: most members watched rather than posted, and when the founders eased off for a single week, new activity fell off a cliff. That’s not a comfortable finding. It’s a valuable one: it named the real problem (activation and founder-dependency, not demand) before a dollar of reinvestment was committed to the wrong thing.
Throughout, we closed the loop out loud, showing members that their feedback changed the format the following week. In a young community, that visible responsiveness is the trust mechanic that everything else depends on.
The pattern
Compress the validation cycle, instrument the outcome, and let the evidence, not attachment, decide what’s worth continuing.
The outcome
In 30 days, gitbent drew 35% activation of non-expert AI builders, validating that the room is real. The experiment surfaced two distinct builder segments and a credible path toward an expert-access model, and, most importantly, produced a measurement framework that turned the continue-or-shutter decision into an evidence-based one rather than an emotional one. The honest read that acquisition succeeded while activation lagged became the single most useful output of all: it pointed the next six months at the right problem.
This is the work I’m drawn to: taking an ambiguous bet and building the discovery and measurement scaffolding that de-risks it, so that whether the answer is keep going or walk away, it’s the right call, made early, for reasons you can point to. gitbent was 30 days and a small room. The discipline behind it is the same one I bring to far larger investments: know what you’re testing, measure what actually matters, and make sure the decision that follows is one the data can stand behind.