Build an Anti-Slop Design Skill with Grok Bot: Ban List + Forced Variety + Independent Review

Input a design brief → run taste rules and variety checks → output AI images with personal distinctiveness

Who

Individual creators or small teams who batch-generate social images and want to avoid the generic 'AI look'

You get

Use a structured ban list, reference library, forced-variety log, independent review, and editor pass to keep AI-generated images stylistically varied and personally recognizable

You need

Grok Bot

Steps

1. Build the ban list and close the loophole

   Write specific visual and copy prohibitions into the bot's design skill. Visual: no purple gradients, glows, or frosted glass; no single default sans for everything; no pill badge above the headline; no icon in a rounded square above every heading; no three identical cards in a row; no stock photos, robots, or glowing brains; no fake dashboard card as the default visual; no highlight box on every headline. Copy: no em dashes, no 'seamless,' no 'It's not just X. It's Y.' Then add the critical loophole rule: if a design choice is the popular first pick for this kind of brief, take the second or third option and say why.

   Execution prompt · Source text。Grok Bot skill / design workflow ban-list and loophole rule node

   Copy prompt

   The ban list
   The list is specific on purpose. No purple gradients, glows, or frosted glass. No single default sans used for everything. No pill badge above the headline. No icon in a rounded square above every heading. No three identical cards in a row. No stock photos, robots, or glowing brains. No fake dashboard card as the default visual. No highlight box on every headline. The copy has its own list: no em dashes, no "seamless," no "It's not just X. It's Y."
   2. The loophole
   A ban list alone fails. The model swaps one banned default for the next most common one. Ban purple gradients and you get a beige background with an italic serif headline, which is now on the list too.
   So the rule under the list matters more than the list:
   If your first pick is the popular choice for this kind of brief, take the second or third option and say why.

2. Create a reference library and write a brief with POV

   Build a reference library of publicly shared, award-winning work; each entry notes what's good about layout, type, color, and spacing. Rule: for a social image, at least one reference must come from outside the UI-card family—poster, editorial page, sign, photo, or chart. Before any image is built, the bot must write three sentences: who this is for, how it should feel, and one non-tech reference (magazine, shop, poster). Later addition: what this piece believes, and who it wants to be to the viewer. If it can't write those, it stops and asks the author. Then a short layout plan, one typeface pairing and why, and one dominant color with at most one accent. Still no building.

3. Enforce variety and log recent designs

   Every new image must differ from each of the last 3–5 images on at least three of these five dimensions: ① palette and background; ② typeface family; ③ how one word or number gets emphasized; ④ layout; ⑤ image type (photo, chart, poster, annotated document, etc.). Before building, the bot writes its picks into a recent-designs log: palette, type, emphasis, layout, image type, and quiet or loud. The next image is checked against that log, not bot memory. The author maintains a short approved rotation: six palettes and six font pairings. Quiet by default: paper backgrounds, serif or mono type, real photos and screenshots first, one light emphasis mark. Loud is capped at about one post in four, never two in a row. Test: Does this look like a person sharing what he does, or a brand account? Brand account fails.

4. Separate review: never let the builder grade its own work

   When the image is done, a different bot or fresh pass reviews it. The reviewer sees only the brief, ban list, references, and finished image—not the builder's reasoning. Four hard-fail checks: ① Sameness—at thumbnail size next to recent images, could it pass for the same template? ② AI look—gradients/glow, everything centered in rounded cards, filler chrome like fake IDs and status pills, or a layout that could belong to any brand; the reviewer must name the one thing that makes it unmistakably the author's, or it fails. ③ Register—too loud, over the cap, or reads like a brand account. ④ Image quality—checked at full size and chat-preview size; blurry, pixelated, blank, or unfinished fails; purposeful white space is fine. Sameness, AI-look, or register failures go back to the variety step for a new combination; small tweaks don't count as a fix.

5. Editor pass and section-by-section fixes

   Passing review only means nothing is wrong. A fresh pass, never the builder, uses the piece the way a real person would on their actual device. It answers four questions: Does it solve the actual problem? Does it hang together? What's missing? Could it belong to anyone? It must name at least three specific detail fixes. Every element must be a choice—if a color or label could be swapped with no loss, it gets decided on purpose or cut. When the direction works but parts still miss, the bot fixes one section at a time (header, background, one chart), re-renders, confirms that part works, and leaves the parts that already work alone.

6. Use real artifacts, redact, and final human approval

   Real screenshots and photos come first. Before anything public, names, account digits, addresses, balances, and health numbers get blurred or swapped for demo data. Then the author reviews it, and nothing posts until the author manually posts it.

The ban list
The list is specific on purpose. No purple gradients, glows, or frosted glass. No single default sans used for everything. No pill badge above the headline. No icon in a rounded square above every heading. No three identical cards in a row. No stock photos, robots, or glowing brains. No fake dashboard card as the default visual. No highlight box on every headline. The copy has its own list: no em dashes, no "seamless," no "It's not just X. It's Y."
2. The loophole
A ban list alone fails. The model swaps one banned default for the next most common one. Ban purple gradients and you get a beige background with an italic serif headline, which is now on the list too.
So the rule under the list matters more than the list:
If your first pick is the popular choice for this kind of brief, take the second or third option and say why.