Project Close

Step 1: Verify the Project State

Before closing, verify all artifacts exist. Direct Claude to list the project files and confirm you have: the PRD (with actual results), evaluation design document, preprocessing decisions document, evaluation results documentation, trained model, serving endpoint code, and MLflow experiment logs.

If anything is missing, go back and produce it. A project close with missing artifacts is not a close.

Step 2: Write the README

Direct Claude to write a project README. It should describe: what was built (an improved churn prediction system for Tunde Mobile focusing on the prepaid segment gap), the approach (artifact creation pipeline -- PRD, evaluation design, documented preprocessing, experiment tracking, per-segment evaluation), and how to run the system (start the FastAPI endpoint, query with curl).

Review what Claude writes. The README is for someone who opens this repository without context. They should understand the project, the approach, and how to use the result.

Step 3: Final Commit

Commit all project files with a clear commit message. Push to GitHub. The repository should tell the full story: the PRD shows what was planned, the evaluation design shows how success was defined, the preprocessing decisions show what choices were made, the experiment logs show what was tried, and the evaluation results show what worked.

Step 4: Reflect on the Pipeline

You just walked through the artifact creation pipeline for the first time: brief intake, PRD, evaluation design, data preparation, training, evaluation, serving, documentation. Each step produced something. The order mattered -- evaluation design before training means you defined "success" before you had results to rationalize, and the PRD before building means you had requirements to build against.

This pipeline is what you'll use going forward. The terrain changes -- different data, different models, different clients. The pipeline stays.