←

Hypothesis-Driven Data Investigation

My BI GPT project let me ask questions in plain English and get back perfect SQL queries, but I was still stuck on the thinking problem. I'd look at our dashboards, notice something interesting, then lose momentum switching between tools to investigate. I needed a way to go from "that's weird" to "here's what the data actually shows" without breaking my train of thought.
Building on that foundation, I created a hypothesis-driven workflow where each idea starts as a conversation with Claude, then gets marked as supported, contradicted, or inconclusive after investigation.
Next.jsTypeScriptPrismaSQLiteClaude Code
>Load up the hypothesis
⏵⏵ ready(click Execute to begin)
I run the app locally and open Claude Code in my project folder. When I load a CSV and type a hypothesis like "template users churn faster," Claude has access to the app's codebase.
Every pattern gets verified through actual SQL queries but instead, Claude writes queries, SQLite runs them, real numbers come back. When we test if template users churn faster, we're not discussing it theoretically. We run the actual queries, get actual counts and calculate percentages.
Claude isn't just generating code in a vacuum. It's working with my local database, using Prisma to write queries, accessing the exact libraries I have installed. The conversation guides the investigation, but the database does the math.
There's almost no UI. Want to load data? Type "load analytics.csv". Need to join tables? Just ask. Since Claude Code can see and run everything in the project folder, I didn't need to build any custodial features.

Let's build something together

>name:
>email:
>msg:
If you'd prefer to email me instead, reach out to hello@davidcibis.com