Build a corpus of what you must never look like

Reverse twenty examples of the thing your brand should never be, then turn what they share into a negative prompt you reuse everywhere.

Everyone can name the look they want. Almost nobody can specify the look they are trying to avoid, which is exactly why it keeps turning up in the work. Twenty anti-references make it concrete enough to exclude on purpose.

Why this will not look generic: A negative corpus is the only artifact here that specifies what to exclude, and exclusion is the half of a brief that normally goes unwritten and therefore unenforced.

What you walk away with

• A prompt for each thing you do not want to look like
• The traits they share, which is the actual thing to avoid
• A reusable negative prompt, saved once and applied everywhere

Typical run: 15 to 30. Below about fifteen you are describing your taste rather than finding a pattern.

This one happens in one conversation

You bring: Fifteen to thirty examples of work in your category that you would be unhappy to be mistaken for.

Open with: Read every prompt in my 'never this' folder. What do they have in common, and write me a negative prompt that excludes it without also excluding things I might legitimately want.

What the thread holds on to:

• The whole anti-corpus, so the shared trait is visible rather than a gut feeling
• The distinction it drew between what is genuinely bad and what is merely common
• The draft negative, which you can tighten by arguing with it

Then you can just ask:

• Which of these traits would I actually want occasionally?
• Is there anything in my own library that breaks this rule?
• Make it shorter; I want something I will actually paste every time

What the reverse actually gives you

Point it at: Twenty pieces of work you would hate to be mistaken for.

What comes back (representative example): The reverses come back with a negative prompt alongside each description, and read together the overlap is specific. Seventeen of twenty centre the subject with symmetrical framing. Sixteen use an even frontal light with no discernible direction, which is what reads as cheap more than any single other choice. Fourteen stack three or more pieces of text at competing sizes. Thirteen use a gradient background rather than a flat or textured one. None of these is objectionable alone, which is why the look survives review one asset at a time and only becomes obvious across twenty.

Things you would never have thought to type:

• even frontal light with no direction, in 16 of 20
• three or more text sizes competing, in 14
• gradient backgrounds rather than flat or textured
• none of it objectionable alone, which is why it survives review

What you change: Nothing about any single image. What you get is a negative you paste into every brief from now on, which is the part of a style guide that never gets written.

How it works

1. Collect what you do not want (In Chrome) - Right-click them as you come across them. This folder is easier to fill than the aspirational one.
2. File them, negative prompts and all - Each reverse returns a negative prompt alongside the description, at no extra cost. Save the set into one folder.
3. Find the shared trait and write the exclusion (In Claude) - Ask it to avoid over-excluding; a negative that bans everything is as useless as none.

Cost and time

This workflow runs at 2 reverses + about 1 minute. Your result carries automatically from each step into the next, so there is no re-uploading between tools. Reverses are billed from your reverse meter and generation steps from your AI Credits, both bundled in every pay-as-you-go credit pack.

Browse all guided workflows or all individual tools. Pay as you go, no subscription, AI Credits never expire.