Source: skill files (frontend-design, taste-skill, soft-skill, ui-ux-pro-max, critique, audit); § Beyond UI from raw/How_to_De-Slop_Every_AI_Output_Forever_With_1_Skill.md and raw/8_Claude_Skills_I_Can_t_Live_Without.md (Ben AI, added 2026-10-28)
A consolidated reference of every banned pattern across all 23 design skills, with the specific alternative to use instead. AI-generated UI has recognizable fingerprints — predictable color choices, default fonts, symmetrical layouts, and decorative patterns that real designers do not use. This guide is the antidote. Pin it. Reference it before every UI review.
Visual and Color
Banned Pattern
Why It Fails
Do This Instead
Purple-to-blue gradients
The #1 AI aesthetic fingerprint
Intentional palettes from ui-ux-pro-max database
Cyan on dark backgrounds
Instantly reads as “AI dashboard”
Tinted neutrals with a single warm or cool accent
Neon glows / neon accents
Cheap sci-fi aesthetic, overused by AI
Subtle accent colors at full opacity --- no glow
Pure black (#000)
Harsh, unnatural on screens
Tinted near-black (#0a0a0a, 1a1a1a with warm/cool cast)
The tables above cover UI. Ben’s de-slop skill applies the same idea to every AI output a person or team sends out: emails, LinkedIn posts, newsletters, proposals, support replies, internal docs. He built it for two reasons. First, his own quality bar drops when he is tired or rushed, and the skill holds the standard you set once (“a spell check, but for slop”). Second, each employee has a different quality bar and their AI agrees with them, which he connects to a16z’s “institutional AI vs individual AI” article (unread here) and the term “workslop.”
Two check layers. A universal layer catches widely agreed tells: em dashes, AI writing patterns and vocabulary, factual errors (claims checked against authoritative web sources), contradictions and unreadable writing. A company layer checks your own standard: tone of voice, brand identity (and visual guidelines for proposals), alignment with strategy and company facts, and whether the asset does its job.
Process. Classify the output type (marketing post, support reply, internal doc), pick the checks that apply, and read one reference file per check (an AI-writing-tells guide, a fact-check method, a voice guide with rules such as “no guru energy” and “no bold claims without backing”). Then three subagents grade the output separately, for less bias, and the skill returns one of three verdicts: good to go, good to go but consider these fixes, or not ready. Each issue comes with its source file and a suggested fix. You choose which fixes to apply.
Demos. A post-call follow-up email got “good to go with fixes”: some AI-writing signs, a claim that didn’t match the call transcript, and small readability and voice issues. A LinkedIn post got “not ready” for breaking hard rules. He also ran it on a newsletter and a sales proposal.
Wire it in. Make it the last step of writing skills, make it a team rule before anything AI-made goes out, and log every run (to the second brain or the skill folder) so recurring failure patterns across the business show up.
Scope. Ben says it is a detector, not a perfect fixer: it lacks the context for every possible asset type, but it makes the team aware of the bar. The free version has blanks for your voice and brand; a paid builder skill walks you through filling them.
Lighter options from the same creator: chain humanizer as the last step of every copy skill (“Can you update my LinkedIn post skill to always run the Humanizer skill after it’s generated the initial output?”), and run a fact-checker skill on drafts, on your own outlines before publishing, and on other people’s articles. See Ben AI’s Claude Toolkit.
To build the company layer quickly, write one reference file per standard (voice, brand, facts) and have the grader cite the file each issue breaks.
Before the gate (added 2026-11-09). The de-slop skill catches slop after drafting. Deslopify Your AI (Nufar Gaspar on The AI Daily Brief, raw/How_to_Deslopify_Your_AI.md) covers the steps before it: state your own view and confidence before prompting (“I’m about 60% sure the problem is onboarding. Push back hard before you agree with me”), ask for several directions instead of one, fix review checkpoints in advance, and run a persona review board. Gaspar calls humanizers “slop trying not to look like slop”, because “humanized text is still flagged about three-quarters of the time” by AI detectors.
Key Takeaways
The “AI color palette” (cyan + purple + neon on dark) is the single biggest giveaway. Eliminate it.
Font choice matters more than any other single decision. Switching from Inter to an intentional font transforms the feel.
Three equal-width cards in a row is the most common AI layout. Use asymmetry.
Content quality kills AI slop faster than visual polish. Real names, realistic numbers, specific copy.
Every banned pattern has a specific alternative --- this is not about avoiding things, it is about choosing better things.
Print this list and check it before every UI review until the patterns become instinct.
For writing and other non-UI output, run a two-layer gate (universal tells plus your company’s own standard) before anything goes out, graded by separate subagents with a clear verdict (Ben AI’s de-slop skill, § Beyond UI).
unslop-ui — the data-driven (frequency-weighted from ~3.2M Reddit posts) detection + CI-failable scanner counterpart to this prescriptive, hand-curated guide.
Hallmark — a Claude Code skill built explicitly as an “anti-AI-slop design skill,” with its own 65-gate generative refusal layer alongside this guide’s hand-curated pattern list.
Walk through each table in this guide and tally how many banned patterns appear
For each match, note the “Do This Instead” column
Rebuild with explicit instructions to Claude: “Use Geist font, not Inter. Use asymmetric layout, not equal cards. Use warm neutrals, not gray. Use ease-out-quart, not bounce.”
Compare the before and after --- each eliminated pattern compounds into a dramatically different result
For copy, take your last AI-drafted email or post and ask Claude to check it in two passes: universal AI-writing tells and factual claims, then your own voice and brand rules, with a verdict of good to go / fix first / not ready (the § Beyond UI pattern)