The Brief and the Plan

Step 1: Read Emeka's Email

Open materials/emeka-followup.md. Emeka is back with an update on the churn model you built in P1. His retention team has been using the ranked predictions weekly. But he's noticed something: the model catches postpaid customers who are about to leave but misses the prepaid ones.

Prepaid subscribers at Tunde Mobile behave differently from postpaid ones. Their top-up patterns are erratic. Their tenure is shorter. When they decide to leave, they just stop topping up and disappear. The model was never designed to distinguish between these two groups.

Emeka also wants structure this time. His board approved the data science budget based on P1's results, and now they want to see methodology. Requirements, evaluation plan, documented results.

Step 2: Talk to Emeka

Open the chat with Emeka. His email gives you the problem, but there are details he hasn't shared. Ask him about the prepaid customer behavior and what his board specifically needs to see.

When you ask about prepaid customers, he'll explain: the postpaid ones complain before they leave -- calls to support, plan downgrades. His team can see it coming. The prepaid ones just stop topping up and disappear. By the time anyone notices, they've already gone.

When you ask about the board, he'll tell you they want documented methodology -- not just "it works" but a structured approach they can review. He'd rather be fixing the model than writing reports, but he knows the documentation matters.

Step 3: Set Up the Project

Open your terminal and start Claude Code:

cd ~/dev
claude

Paste this prompt:

Create the folder ~/dev/ml/p2. Download the project materials from https://learnbydirectingai.dev/materials/ml/p2/materials.zip and extract them into that folder. Read CLAUDE.md -- it's the project governance file.

Claude will download the materials and set up the workspace. After it finishes, look at what's in materials/. You should see CLAUDE.md, emeka-followup.md, prd-template.md, subscribers-v2.csv, data-dictionary-v2.md, and tickets.md.

Open materials/tickets.md and scan the full ticket list. This is the work breakdown for the entire project. The tickets give structure, but the decisions within them are yours.

Step 4: Profile the Updated Dataset

Direct Claude to load materials/subscribers-v2.csv and profile it. You want the same basics as P1 -- row count, column types, summary statistics, missing values, churn distribution -- plus something new: the breakdown by segment.

Something like: "Load subscribers-v2.csv and give me a full profile. Include the churn rate overall and broken down by the segment column (prepaid vs postpaid)."

Open materials/data-dictionary-v2.md and compare what Claude reports against the dictionary. Notice the segment column -- it's new. Emeka's data team added it after they noticed the prepaid gap.

Look at the numbers. Overall churn is around 8%, same as P1. But break it by segment: prepaid customers churn at roughly 12%, postpaid at roughly 4%. The overall number hides a three-to-one gap between segments. A model evaluated on overall metrics could look fine while completely failing on the prepaid customers Emeka needs to catch.

Step 5: Write the PRD

Open materials/prd-template.md. This is a requirements document template with section headings and prompts. The PRD captures what you're building, why, and how you'll know it works -- before you build anything.

Direct Claude to draft a PRD using this template. Feed it what you know from Emeka's email and your conversation: the prepaid gap, the board's need for documented methodology, the evaluation requirements. Review what Claude produces. The Problem Statement should name the prepaid churn gap specifically. The Success Criteria should include metrics with rationale -- not just "use F1" but why F1 matters for this problem. The Evaluation Strategy should describe per-segment evaluation.

This is your first time writing a PRD. The document is the contract between what Emeka asked for and what you'll deliver. Get it right now and every subsequent decision has something to anchor to.