Data Cleaning and Structuring of Messy Client Exports

The file-upload and code-execution feature needs a paid ChatGPT plan. The free tier allows limited uploads, not the sustained back-and-forth this workflow needs.

1. Go to chatgpt.com and sign in.
2. Click your profile icon in the corner and check your plan under Settings.
3. If you're on the free tier, upgrade to Plus before continuing.

What you should see: A settings panel showing your current plan name and billing details. Troubleshooting: If your firm has an enterprise ChatGPT account, ask your IT or data governance contact whether client files are approved to upload there before you use a personal account instead.

1. Click New chat in the sidebar.
2. Click the paperclip or + icon in the message box.
3. Select Upload from computer and choose your file.

What you should see: A file preview chip appears above the text box, showing the filename and size. Troubleshooting: Very large files (tens of thousands of rows) may take a minute to process. If the upload stalls, split the file into smaller sheets and upload the worst-offending one first.

Don't just say "clean this." Tell the model what the output should look like, since operations research work usually feeds a specific downstream model.

1. In the message box, describe your source columns and what the cleaned version needs to look like: consistent types, a single date format, no duplicate keys.
2. Send the message.

What you should see: The model responds with a short summary of what it noticed in the file (data types, obvious duplicates, missing values) before writing any code.

1. Ask directly for the cleaning pass, and require a change log: "Clean this dataset to match the structure I described. Run the code, then give me a summary of every row you changed or removed, and why."
2. Let it write and execute the code. This happens in a sandboxed environment, so you'll see it work through the problem, not just receive a final answer with no explanation.

What you should see: A written explanation of the cleaning logic, followed by a downloadable file link and a table or list summarizing what changed. Troubleshooting: If the code errors out, it's usually a data type problem (text in a numeric column) or a date format the model didn't anticipate. Paste the error back and ask it to handle that specific case.

1. Click the download link to get the cleaned CSV or Excel file.
2. Open it alongside the original and check 10-15 rows by hand, especially any the model flagged as altered.

What you should see: A file that matches your requested structure, with the change summary matching what you observe in the actual rows.

If a "duplicate" removal or a filled-in blank looks wrong, say so directly: "Row 214 wasn't actually a duplicate, it has a different project code. Restore it and re-check for that pattern elsewhere." The model will revise and rerun.