How to Create a Test Data Plan with AI (2026): Fake Rows, Real Rules

A test data plan written with AI is a page that says which fields a test needs, the rules those fields must follow, where the rows may live, and the ban on copying production. The model may draft fake values that follow rules you wrote. It may not paste a real customer, a real card, or a real address. The place a test must not touch production is what an AI tool sandbox is. The rows you re-run to judge a model, which is a different set, are what an AI eval dataset is.

The checklist that says how a migration is checked is how to write a migration runbook with AI. The undo if a test writes somewhere it should not is how to create a rollback plan with AI.

Prices are US list prices excluding tax from directory pages checked October 16, 2026. Published October 24, 2026. We did not generate a sample database. This page is not a privacy policy.

What fits in the plan

Block	Who owns it	The model may
Fields	The columns you named	List them. It may not add a column of real-looking government IDs
Rules	Formats and relationships you wrote	Apply them to fake rows. It may not relax them
Volume	A count you chose	Repeat up to that count. It may not “make it realistic” by copying production
Home	The sandbox or the fixture file	Name it. If the home is production, stop
Ban	What must never appear	Copy the ban. A row that violates it is a defect

flowchart TD
  rules["You paste fields, rules, volume, and the ban on production data"] --> home{"Is the destination a sandbox or a fixture, not production?"}
  home -->|No| stop["Write Destination unsafe and do not generate rows"]
  home -->|Yes| draft["Claude drafts fake rows that follow the rules"]
  draft --> scan["You scan for any value you recognize from production"]
  scan --> keep["Keep the fixture. Discard the chat if it wandered"]

A fake email at a domain you own is a rule. A fake email at a real customer’s domain is a leak. If you cannot tell, do not generate the column. Leave it blank and say the rule is missing.

Where the plan is written

The draft. Claude Pro is $20 a month, or $17 billed annually, when the field list is long. Use a business-appropriate account if the rules themselves describe sensitive fields. The review is Claude review.

The fixture in the repo. Cursor Plus is $8 a month, or $80 a year, when the rows become a file. You read the diff. A generated row that includes a phone number you did not ask for comes out. The ladder is Cursor pricing explained.

A small set. ChatGPT Plus at $20 a month can draft a dozen rows. Business, at $20 a user a month billed annually, is the row when the schema is confidential. Do not paste a production export “as an example.”

The page. Notion AI Business is $20 a seat a month if the plan is what QA opens. The fixture file is still the data. The page points at it.

A prompt that holds the line:

“Draft fake test rows for the fields and rules I pasted, up to the count I named. Do not use real people, real emails, or real account numbers. If I did not name a non-production destination, write Destination missing and stop.”

What not to delegate

Do not ask the model to anonymize a production dump inside a chat. That dump is already in the prompt. Do not commit fixtures that fail the ban. Do not point a test at production because the sandbox was down. The quality set you use to judge answers, if this plan is about a model rather than an app database, stays on the eval-dataset page so the two files do not get mixed.

FAQ

Can I start from a production sample if I delete the names?

Not in the chat, and not as the default. A deleted column can come back in a free-text field. Build the rows from rules. If your security team has an approved masking pipeline, that pipeline is the path, and this prompt is not a substitute for it.

How many rows?

The count that exercises the rule: a blank, a maximum, a duplicate, a row that should fail. A million rows is a generation job with its own review. The plan names the count before the model runs.

Where may the file live?

In the repo if it contains nothing real, or in the sandbox’s fixture store. It may not live in a ticket, a slide, or a personal drive. The plan names the home.

Who throws the file away?

The owner, on a date the plan includes, if the fixture was temporary. A test file that sits for a year becomes something people treat as real. Date it.

Last verified

October 16, 2026 for Claude, Cursor, ChatGPT, and Notion AI. Published October 24, 2026.

Sources: Claude pricing, Cursor pricing, ChatGPT pricing, Notion pricing.

Published by Andrew for AI Tools Optimizer.

Cursor is in the coding directory. Directory pages may use affiliate links. A purchase through them may earn us a commission at no extra cost to you.