Prompt Atlas — fictional Harbor Thread sales dashboard test
Updated October 11, 2026. All data are fictional.

Paste initial-prompt.txt into ChatGPT. It contains sales-v1.csv. original-dashboard.html has a documented duplicate-row bug. corrected-dashboard.html matched the original CSV in All/Store/Online preview checks.

Original reference: gross $660, refunds $60, net $600; 6 lines, 5 orders. Store net $350, Online $250.

Local follow-up: John Deris opened corrected-dashboard.html on his computer, uploaded the fixtures, and copied dashboard output into the test record. No local screenshots were verified.

sales-v2.csv upload: gross $720, refunds $60, net $660; 7 lines, 6 orders. Store $350, Online $310. Copied results matched the independent reference.

sales-invalid.csv starts from v1; L01/A01 gross $200 is replaced with not-a-number. Row 2 was excluded as Invalid required value. Remaining output: gross $460, refunds $60, net $400; 5 lines, 5 orders. A01's Belt line remained. This is a partial total: the audit explains the exclusion, but copied output did not show a prominent KPI warning. The whole file was not rejected.

Recovery: reloading sales-v2.csv restored $660 net, 7 lines, 6 orders. Product net: Jacket $300, Belt $90, Jeans $180, Shirt $90. Audit returned to 5 exclusions.

Inspect generated HTML before use. This small test does not establish arbitrary-file accuracy, large-file performance, repeat-run reliability, live integrations, or business outcomes.

Prompt formatting: CSV line breaks restored from the fixture; recorded wording preserved.
