How to Build a Weekly Data Analyst with Claude Cowork

Download the files into a local working folder

Download the files from the Novamart Google Drive folder into a local folder. Use that folder as your working directory for the project.

Create the first briefing file for Claude Cowork

Ask Claude Cowork to create a claude.md file in the project folder with the company, project, task, and input-file details:

Create a claude.md file within this project folder with the following information —
Company and project: Novamart; project is the weekly analysis automation
The task is to run a weekly analysis of Novamart’s ecommerce performance and produce a report and summary for leadership
You’re building a personal AI data analyst that answers a standing list of weekly questions, makes supporting charts, and delivers by email and saves a report locally.
This will be run automatically, on a command or schedule
The data: all data files are in the folder novamart-data
events.csv
reviews.csv
products.csv

Create and check the DuckDB database

Ask Claude Cowork to create a local DuckDB database named novamart.db in the main project folder, using the files in novamart-data. Have it verify the setup with basic row-count, column-name, and missing-value checks.

Create a local database using DuckDB called novamart.db in the main folder, using all files in the novamart-data folder. Check that the setup works by perform simple checks - #rows in each file, column names, count of missing values.

Add the file and database details to the briefing

Once the database checks are complete, ask Claude Cowork to record the file columns and database information in claude.md.

Add the info about the files columns and the database into the briefing file (claude.md)

Ask the first analysis question

Use the database to answer the initial question about the six-week conversion trend:

Answer the following question from the database -
What is the trend of conversion for the past 6 weeks?

Define the metrics and customer-rating source

Add the metric definitions to claude.md so the analysis uses consistent calculations and time windows.

Use the following definitions and add them to the claude.md file -
## View-to-purchase rate aka “conversion”
Purchase events/ view events, same product_id, same time window.
## Cart-to-purchase rate aka “card addition rate”
purchase events / cart events, same product_id, same time window.
## "Last week"
Monday 00:00:00 to Sunday 23:59:59, most recent complete calendar week.

Also make the source of customer ratings explicit:

Add the following to claude.md file -
Customer ratings always come from the reviews table (submitted by purchasers after a
transaction). Never use products.rating for customer satisfaction - that's catalogue
metadata, not customer feedback.

Then rerun the conversion-trend question and explicitly tell Claude Cowork to use the definition in the briefing file:

What is the trend of conversion for the past 6 weeks? Use the definition from the claude.md file.

Create a chart that states the finding

Ask Claude Cowork to create a chart with Python’s Matplotlib. The chart title should communicate the finding, not merely name the metric or topic. For example, use “Conversion Is Highest for Electronics” rather than “Conversion Rate by Category”.

Create a chart using Python matplotlib. Titles state the finding, not the topic: e.g. "Conversion Is Highest for Electronics", not  "Conversion Rate by Category"

Answer the weekly performance questions

Ask Claude Cowork to answer the following questions from the database and include a chart or other evidence for each answer:

Answer the following questions -
What is the overall conversion rate last week vs. previous week?
Which product categories have the highest and lowest conversion rates?
Which product categories got the worst reviews last week?
Add chart/evidence for each question.

Create the weekly leadership summary

Ask Claude Cowork to turn the findings into a Word document. Keep the summary concise, tie each headline to its evidence, and avoid guessing at causes that the data does not establish.

Create a summary doc using the answers above. Executive summary - 2-3 sentences on what's true across all findings, no hedging. Each finding: a bold headline sentence, its chart/table, and a recommended action only if the finding actually warrants one. If a cause isn't confirmed by the data, recommend investigating it - never guess at one.
Save as a Word doc named weekly-insight-[date].docx in a reports/ folder.

Connect Gmail in Claude Cowork

In Claude Cowork, open Settings → Connectors/Integrations, add Gmail, sign in, and allow the connection.

Send the weekly summary email

Request an email with only the short summary of the week’s most important findings. The instructions below also specify not to attach the full Word report and to save a fresh copy each time.

Send me an email summary with the 2-3 sentence summary of the week’s most important findings. Never attach the full docx file. Always save a fresh copy, even if one exists for the week

Create a skill for the complete workflow

Finally, ask Claude Cowork to package the process as a skill, covering the workflow from database creation through the email step.

Create a skill for the entire workflow, starting from creating the database from the input folder, to the email step. Add the skill markdown file to this project folder.