AI Sanitize

Find and remove the patterns that make work read as machine-made. The patterns are catalogued in three references, one per medium. This file is the procedure.

The rule behind every reference

AI output combines familiar patterns without a reason for choosing them. No single pattern proves anything. Inter is not an AI font, purple is not an AI color, and an emdash is not an AI character. People used all of them first.

What reads as generated is the combination of patterns plus the absence of intent. So judge every candidate with two questions:

1. Removal test. If you delete it, does the reader lose information or the user lose an ability? If not, it is decoration without intent. Remove it.
2. Reason test. Can you name why this choice fits this product, audience, or argument? "It looks modern" and "it sounds polished" are not reasons. If there is a reason, keep it, even if it appears in the catalog.

A page with one gradient is fine. A page with a gradient hero, an eyebrow, a huge H1 over a gray subtitle, pill buttons, glass cards, purple accents, a pulsing badge, Inter, JetBrains Mono, and "Elevate your workflow" is the tell.

Modes

• Edit (default). Change the files or text, then report.
• Report. The user says "report", "audit", "just list", or "don't change anything". List tells with location and proposed fix. Edit nothing.