Sharing my prompt template for generating A/B test hypotheses from past data.

Hey everyone. I've been knee-deep in comparing CRM reporting dashboards again (surprise, surprise) and it got me thinking about how we use AI for analysis. We all have these past sales cycles, support tickets, or marketing campaigns sitting in our CRMs—goldmines of data that just... sit there.

I've been using DeepSeek Chat to help me generate concrete, actionable A/B test ideas from that historical data. After a lot of tweaking, I've landed on a prompt template that works pretty well across different platforms (I've tried it with Salesforce, HubSpot, and Pipedrive data exports). The key is to force it beyond vague suggestions like "test subject lines" and get to measurable hypotheses.

Here's my current template structure:

**Role & Goal:**
"You are a data-driven growth analyst. My goal is to generate high-potential A/B test hypotheses from a dataset of past [campaigns/sales calls/support interactions]. Focus on identifying clear independent variables (what we change) and dependent variables (what we measure)."

**Data Context:**
"I will provide you with [describe data, e.g., 'a CSV with columns for: email subject line, content type, send day, open rate, click-through rate, conversion rate']."

**Output Instructions:**
- First, identify 2-3 patterns or anomalies in the data you'd want to explore.
- For each, generate a specific, falsifiable hypothesis using this format: "Changing [independent variable] from [current typical state] to [proposed variation] will result in a [predicted change] in [dependent variable]."
- Prioritize hypotheses where the variable is easily actionable by my team.
- Note any potential confounding variables we'd need to control for.

**Example:**
If I fed it my last email campaign data, a good output might be: "Changing the email send day from Tuesday (current) to Thursday (proposed) will result in a 10% increase in click-through rate. *Control for: content type, subject line format.*"

This forces a level of specificity that's been super helpful. It's not perfect—you still need to sanity-check the logic—but it's a fantastic starting point. I find it works better if you paste in a small, clean sample of your actual data rows.

Has anyone else been using DeepSeek for this kind of analysis prep? I'm curious if you've found ways to refine it further, especially for pipeline or churn data.