Skip to content
Back to search

deslop

SkillWorks with Claude CodeGitHub

Anti-AI-slop rules and audit for any UI work. Use before and after building or changing anything visual (web pages, apps, dashboards, artifacts, slides, emails, video graphics, brand work), and whenever asked to audit, review, "de-slop", "unslop" or check a design or project for AI slop or a generic, templated, "vibe-coded" look. Gives a house style, a subtract-first clean-up pass, and a scanner that lists slop patterns by file and line. Reads your own brand from references/brand.md when that file exists. When an installed design skill conflicts with it, this one wins.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/pratyushjain98--deslop/badge.svg)](/listings/pratyushjain98--deslop)

HTML

<a href="/listings/pratyushjain98--deslop"><img src="/api/listings/pratyushjain98--deslop/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

git clone https://github.com/PratyushJain98/deslop ~/.claude/skills/deslop

Claude Code loads it in the next session. It triggers on UI work and on requests to audit or "de-slop" a design, or you can call it with `/deslop`. To use it in one project only, clone it into that project's `.claude/skills/deslop` instead.

On claude.ai or the Claude desktop app, zip the folder and upload it as a custom skill.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 122 lines
  1. ---
  2. name: deslop
  3. description: Anti-AI-slop rules and audit for any UI work. Use before and after building or changing anything visual (web pages, apps, dashboards, artifacts, slides, emails, video graphics, brand work), and whenever asked to audit, review, "de-slop", "unslop" or check a design or project for AI slop or a generic, templated, "vibe-coded" look. Gives a house style, a subtract-first clean-up pass, and a scanner that lists slop patterns by file and line. Reads your own brand from references/brand.md when that file exists. When an installed design skill conflicts with it, this one wins.
  4. ---
  5. # deslop
  6. AI-made interfaces look alike because every generator reaches for the same defaults. A slop pattern is **a default nobody chose**.
  7. **When you build, avoid every pattern in `references/patterns.md`, all of them, not just the top 25.** The top 25 is only a priority order for fixing existing work. The only exceptions are the choices in House style below and the rules of a brand you have been given.
  8. **When you audit, judge clusters.** A single hit is rarely the problem, and it isn't proof that an AI made the thing. Four or more distinct tells on one page reads as "templated".
  9. Treat this as a standing requirement for UI work, and apply it without being asked.
  10. ## Build mode: whenever you design or change UI
  11. 1. **Load the brand.** If the project has a design system, brand book or tokens, use them. If `references/brand.md` exists in this skill, read it: it holds the user's own brand (logo files, colours, type and house rules), and it wins over the house style wherever they differ. `references/brand.example.md` shows the format. When you review someone else's site, follow "Reviewing someone else's site" below. If there is no brand, apply the house style.
  12. 2. **Set the direction before you draw.** Defaults are what you get when you skip this step. Write five short lines, taken from what the product does and who it's for, never from what's trending:
  13. - **Product:** what it does and for whom, in one plain sentence.
  14. - **Character:** two or three concrete words, such as "exact, calm, a little warm". Never clean, modern, sleek, minimal or premium, because those describe every template.
  15. - **World:** the place or material the product belongs to, such as an exam hall, a ledger or a site office. It shapes the accent, the imagery and the copy. It must not tint the backgrounds: grounds stay white, black and grey.
  16. - **Moments:** the one or two places that get the richness, such as a real photo, a chart drawn with care or a key figure. Everything else stays quiet.
  17. - **Signature:** one small detail only this product has, used consistently, such as a heavy rule taken from the logo's baseline, or progress marks shaped like part of the mark. One detail, not five.
  18. Then fix the tokens: typeface, palette (neutrals plus at most one accent, used for meaning), a 4px spacing grid, a small radius scale (nested corners add up: inner radius = outer radius minus the padding between them), and a few type sizes.
  19. 3. **Use real content.** Real names, figures, dates and screenshots. No invented stats, fake logos, initials-in-circles, "Acme", or lorem-shaped copy. If a fact is missing, leave a visible placeholder such as [PRICE].
  20. 4. **Run the clean-up pass** in `references/checklist.md`: delete, de-box, snap spacing, cut the type scale, strip meaningless colour, rewrite generic copy, swap in real assets, and limit motion. It subtracts first, because every good de-slop starts by removing things.
  21. 5. **Scan it:** `python <this skill>/scripts/audit.py <path>`.
  22. - Fix every **certain** high finding.
  23. - Look at each finding listed under "Check these by eye". Those patterns can misfire, so confirm one before you change anything for it.
  24. - Fix each medium finding, or keep it on purpose (see Ignoring).
  25. 6. **Look at it.** Screenshot it at desktop (about 1280px) and phone (about 390px) width, using whatever browser tool is available. Walk the "Check by eye" list that the audit prints. Static scanning can't see composition, rhythm, centring or whether copy is generic.
  26. 7. **Don't trade one cliché for the next.** The fixes the internet recommended in 2025 are now tells themselves:
  27. - serif display type, italic accent words, cream paper and terracotta
  28. - "distinctive" free fonts (Fraunces, Newsreader, Bricolage, Space Grotesk, Instrument Serif, Playfair)
  29. - monospace labels as costume, 01/02/03 numbering, tiny tracked capitals
  30. - frosted glass, bento grids, grain, aurora blobs
  31. - extreme weight contrast, and staggered fade-ins on every section
  32. ## Changing existing work: decide how much, then ask
  33. Read the audit honestly before planning anything:
  34. - **Heavy slop**, with many tells across colour, type, icons and copy: run the full direction step and the whole clean-up pass.
  35. - **Light slop**, a mostly designed screen with a few tells: polish only. Fix what the audit found, take the direction from the existing brand, and keep its palette, type and layout. Put bigger ideas in the report as suggestions.
  36. - **No slop:** say so plainly and suggest two or three small refinements at most. Never change a good screen just to make it different.
  37. On work that already exists, show the findings and the direction, then wait for a yes before editing, because changing direction later means redoing everything. Skip the wait only when the user has said to just do it. Keep the product: its features, facts and order of sections stay, and nothing new is claimed.
  38. ## Reviewing someone else's site
  39. Apply these standards to their brand. Don't put your own logo, colours or type on their site.
  40. - **Keep theirs:** their logo, their brand colour (as the one accent), their real photos, their content and the order of their sections.
  41. - **Apply these rules:** one typeface at three or four sizes; white, black and grey doing the work, with their colour only as a signal; lines and space instead of cards and boxes; one heading per idea, in sentence case; real figures instead of filler; nothing animating on load; and every pattern in `patterns.md` removed.
  42. - **Describe the improved version concretely.** For example: "the three feature cards become one ruled list", or "the gradient hero becomes their best photo with one heading and one button".
  43. - Any document you write about the review (a report, a proposal) uses your own brand, if `references/brand.md` has one.
  44. ## Audit mode: checking an existing project
  45. 1. Run `python <this skill>/scripts/audit.py <project root>`. Useful options:
  46. - `--json`: machine-readable output.
  47. - `--min-severity medium`: cut the noise.
  48. - `--fail-on high`: exit with code 2 if any certain finding is at or above that severity. Add `--strict` to count the check-by-eye findings too.
  49. - `--hook`: run as a PostToolUse hook that scans each edited file, for Claude Code (Write/Edit) or Codex (apply_patch). `--also-fail ID,ID` makes those rules block whatever their severity. Installing the hook changes the user's settings, so ask first.
  50. - `--url <link>`: audit a live page when you don't have its code. It reads the page and its stylesheets as text. Pages drawn by JavaScript need screenshots as well.
  51. Big apps repeat the same layout on many pages, so group findings by the shared layout or component, not by page.
  52. 2. Screenshot the main screens at both widths and walk the visual checklist.
  53. 3. Report the findings in the conversation, not in a separate file, unless the user asks for one. Lead with the verdict ("templated: N tells" or "clean"). Then list the fixes, highest impact first, each with the file and line and the one-line fix. Note what you could not check, such as motion, rendered colours, or pages behind a login.
  54. 4. Change code only when asked. Then fix the high-impact patterns first, re-run the audit, and re-shoot the screens.
  55. ## House style
  56. This is the default when the project has no brand of its own.
  57. - **Type:** Geist or Outfit for text, Geist Mono possibly for figures and small labels, Geist Pixel only for posters and display. One family per surface, at three or four clearly different sizes. Headings sit at 600 to 650; 800 to 900 is a tell.
  58. - **Never:** serif display; italics as decoration; Fraunces, Newsreader, Bricolage Grotesque, Zalando Sans Expanded, Geist Serif, Instrument Serif, Playfair or Space Grotesk.
  59. - **Colour:** white, black and greys do the work. The one accent colour is a signal, never a surface: links, the live or current item, progress, focus, or one key figure.
  60. - **No:** tinted fills, blue boxes, cream paper, terracotta, purple or indigo, gradients, glows.
  61. - **Structure:** lines and space, not boxes.
  62. - Tables are ruled: a heavy line top and bottom, and a hairline under the header.
  63. - A key point gets a thin line beside unboxed text, never a callout box with a dot and a label.
  64. - One heading per idea. A subheading only when it adds a fact the heading can't hold.
  65. - Labels are in sentence case.
  66. - No dot before a label ("● Active") and no "·" between two figures ("3 new · 2 changed"): space and the figure's own colour do it. A "·" between words in a subtitle or caption is fine.
  67. - **Identity** comes from the brand's own details, never from trends: for example, a heavy rule taken from the logo's baseline, or a shape taken from its mark.
  68. - **Allowed** even though some detectors flag them: Geist, Outfit, ruled tables, and a thin line of 2px or less beside unboxed text.
  69. ## Design skills that conflict
  70. Some popular design skills for Claude Code and Codex push patterns that are now slop:
  71. - **design-taste-frontend (tasteskill):** a staggered entrance on every section with a minimum motion count, code-drawn pictures, tinted grounds, and serif or 800-weight type from its draw. Use `--keep palette,type` and skip its motion floor.
  72. - **motion-framer and gsap-scrolltrigger:** hover-scale, glass cards, fade-in, parallax and stagger in their examples.
  73. - **anatomy:** tiny mono capitals in its figure card.
  74. - **gstack design-consultation:** Fraunces, Instrument Serif, mono for data, grain, a new look every time.
  75. - **gstack design-review and plan-design-review:** gradients instead of flat grounds, and two or three motions on landing pages.
  76. - **compound-engineering frontend-design:** serif headlines, earth tones, stagger, parallax and hover-scale in its examples.
  77. Older skills such as design-taste-frontend-v1, high-end-visual-design, minimalist-ui, stitch-design-taste, gpt-taste and industrial-brutalist-ui push more of the same; `references/patterns.md` lists each one under "Older design skills".
  78. Where any of them disagrees with this skill, follow this one. `references/patterns.md` (section "Design skills that push patterns that are now slop") lists exactly what each one pushes.
  79. ## Ignoring a pattern on purpose
  80. If the brief really wants something the scanner flags, keep it. Mark the line with `deslop-ignore: ID` and give a reason, or mark the whole file with `deslop-ignore-file: ID, ID`. Never suppress a whole file just to get a clean report.
  81. ## Files
  82. - `references/brand.example.md`: the format for your own brand file. Copy it to `references/brand.md` and fill it in.
  83. - `references/checklist.md`: the clean-up pass, in order, with concrete numbers.
  84. - `references/patterns.md`: the full catalogue of 123 patterns. Each has what it looks like, why it reads as AI, how to detect it, the fix, and the evidence. It ends with the top 25 and what static scanning can and can't catch.
  85. - `scripts/audit.py`, `scripts/rules.json`: the scanner. Python standard library only, so it runs on Windows, macOS and Linux with nothing to install.
  86. - `scripts/RULES.md`: one line per rule, generated from rules.json by `scripts/make_rules_md.py`.
  87. - `tests/`: fixtures and `run_tests.py`.
  88. ## Credits
  89. The direction step, the three modes, the dashboard rules, nested corners, and several copy and spacing rules are adapted from **unslop** by Damir Sokolovsky (https://github.com/heydamir/unslop, CC BY 4.0), with changes to fit this skill's house style. unslop's font table, status-dot signatures, mono meta labels, load-in stagger and tinted surfaces were deliberately left out. See NOTICE.
  90. ## Honesty
  91. Say what the scan can't see. A clean scan is not a clean design until the screenshots have been checked too. Never call something slop-free without looking at it.
README.md · README · 142 lines
  1. # deslop
  2. deslop finds the patterns that make AI-built interfaces look alike: purple gradients, a small label above every heading, three identical feature cards with icons, cream paper with an italic serif headline, and more than a hundred others. It is a skill for Claude Code and other coding agents, plus a scanner you can run on its own.
  3. It has three parts:
  4. - **Rules for the agent** (`SKILL.md`). Read before building or changing any UI: set a direction, use real content, run a clean-up pass, then scan and look at the result.
  5. - **A catalogue** (`references/patterns.md`). 123 patterns, each with what it looks like, why it reads as AI-made, how to detect it, the fix, and the sources behind it.
  6. - **A scanner** (`scripts/audit.py`). Lists each pattern it finds by file and line, in HTML, CSS, JS, TS, JSX, TSX, Vue, Svelte, Astro and MDX. Python standard library only.
  7. ## The problem
  8. Code generators reach for the same defaults, so the pages they make look alike: the centred hero with a pill badge, the indigo button with a glow, the row of three cards, the "Ready to get started?" band. People now recognise that look as "vibe-coded" and trust it less.
  9. The usual advice makes it worse. The fixes recommended in 2025 (serif display type, cream backgrounds, a terracotta accent, Fraunces or Space Grotesk, monospace labels, bento grids) became the next set of defaults within a year. deslop treats those as tells too, and gives the agent a short house style to fall back on when a project has no brand.
  10. The scanner can't see composition, motion or whether the copy says anything. So every report ends with a screenshot checklist, and the skill tells the agent to look at the page at desktop and phone width before calling it clean.
  11. ## Install as a Claude Code skill
  12. ```sh
  13. git clone https://github.com/PratyushJain98/deslop ~/.claude/skills/deslop
  14. ```
  15. Claude Code loads it in the next session. It triggers on UI work and on requests to audit or "de-slop" a design, or you can call it with `/deslop`. To use it in one project only, clone it into that project's `.claude/skills/deslop` instead.
  16. On claude.ai or the Claude desktop app, zip the folder and upload it as a custom skill.
  17. ## Use with other agents
  18. `SKILL.md` is plain Markdown, so any agent that can read files can follow it.
  19. - **Codex and other agents that load skills from a folder:** link this folder into that agent's skills folder rather than copying it, for example `ln -s ~/.claude/skills/deslop ~/.agents/skills/deslop`. Check your agent's documentation for the folder it reads.
  20. - **Anything else:** add a line to the project's `AGENTS.md` or system prompt, such as "For any UI work, read `path/to/deslop/SKILL.md` and follow it."
  21. ## Run the scanner on its own
  22. You need Python 3.9 or later and nothing else.
  23. ```sh
  24. python3 scripts/audit.py path/to/project # scan a folder or files
  25. python3 scripts/audit.py src --min-severity medium # hide low-severity findings
  26. python3 scripts/audit.py . --fail-on high # exit 2 if any certain high finding (for CI)
  27. python3 scripts/audit.py . --json > deslop.json # machine-readable output
  28. python3 scripts/audit.py --url https://example.com # audit a live page from its HTML and CSS
  29. python3 scripts/audit.py --list-rules # print the rule table
  30. ```
  31. Folders named `fixtures`, `__fixtures__`, `__mocks__`, `stories` and `.storybook` are skipped unless you pass `--include-examples`. Built output, minified files and files marked as generated are skipped too.
  32. Here is the start of the report for `tests/fixtures/slop.html`, a page written to contain as many tells as possible (`python3 scripts/audit.py tests/fixtures/slop.html --max-locs 2`):
  33. ```text
  34. # deslop audit: tests/fixtures/slop.html
  35. **templated: 69 tells** in `slop.html` (42 certain, 27 to check by eye): BOX-02, BOX-04, BOX-05, BOX-06, CMP-01, CMP-02, CMP-03, CMP-04, CMP-05, CMP-06, CMP-07, CODE-01, CODE-03, CODE-04, COL-01, COL-02, COL-03, COL-04, CPY-01, CPY-02, CPY-03, CPY-04, CPY-05, CPY-06, CPY-07, CPY-08, CPY-10, CPY-11, CPY-13, CPY-14, CPY-16, CPY-17, DAT-02, DAT-03, DAT-04, DAT-05, DAT-07, DEC-01, DEC-02, DEC-03, DEC-04, DEC-05, DEC-06, DEC-07, DEC-08, DEC-09, DEC-13, HDG-01, HDG-02, HDG-03, HDG-05, HDG-07, IMG-03, IMG-06, LAY-11, MOT-01, MOT-04, MOT-05, MOT-06, MOT-07, NAV-01, NAV-02, NAV-03, NAV-04, TYP-01, TYP-02, TYP-05, TYP-06, TYP-10
  36. Scanned 1 file. 50 patterns found for certain and 29 to check by eye. Slop score 178 (lower is better; it weights severity by how many files a pattern spreads across, and counts the check-by-eye ones at half).
  37. ## High (15)
  38. ### CMP-01 Gradient or glowing primary button: 2 hits in 1 file
  39. Why it reads as AI: The gradient 'Get started' button with a coloured glow is the first-wave CTA every scanner lists.
  40. Fix: A flat fill in one colour, with a label that names the action.
  41. - `slop.html:37` .btn-primary { background: linear-gradient(90deg, #6366f1, #a855f7); box-shadow: 0 0 24px rgba(99, 102, 241, .6); border-radius: 16px; }
  42. - `slop.html:77` <button class="bg-gradient-to-r from-indigo-500 to-purple-600 shadow-lg shadow-indigo-500/50 rounded-2xl px-6 py-3 transition-all">Start free</button>
  43. ### CODE-01 Untouched shadcn/ui theme: 3 hits in 1 file
  44. Why it reads as AI: The default shadcn tokens left as installed make the product look like every other one built from the same starter (23.5% of sites in one sample).
  45. Fix: Treat shadcn as structure. Set your own palette, radius and type tokens before building.
  46. - `slop.html:13` --primary: oklch(0.205 0 0);
  47. - `slop.html:14` --radius: 0.625rem;
  48. - and 1 more
  49. ```
  50. The full report goes on to the medium and low findings, then a "Check these by eye" section for patterns that have honest uses, then the screenshot checklist. The same command on `tests/fixtures/clean.html` finds nothing and prints only the checklist.
  51. A page with four or more distinct high or medium tells is reported as templated. One hit on its own is rarely a problem, and it isn't proof that an AI made the page.
  52. Exit codes: `0` when nothing reaches `--fail-on` (or you didn't pass it), `2` when a certain finding does, `1` for a usage or read error. Add `--strict` to let the check-by-eye findings fail the run too.
  53. ### As an editor hook
  54. `--hook` reads a Claude Code or Codex PostToolUse payload on stdin, scans only the files the agent just edited, and exits 2 with the findings on stderr so the agent sees them and fixes them. For Claude Code, add this to `settings.json`:
  55. ```json
  56. {
  57. "hooks": {
  58. "PostToolUse": [
  59. {
  60. "matcher": "Write|Edit|MultiEdit",
  61. "hooks": [
  62. { "type": "command", "command": "python3 ~/.claude/skills/deslop/scripts/audit.py --hook --fail-on high" }
  63. ]
  64. }
  65. ]
  66. }
  67. }
  68. ```
  69. `--also-fail CPY-05,DEC-13` makes particular rules block whatever their severity.
  70. ## Keeping a pattern on purpose
  71. If a brief really calls for something the scanner flags, keep it and say why in a comment. Any comment syntax works.
  72. ```html
  73. <p class="eyebrow">Case study</p> <!-- deslop-ignore: HDG-01 the client's print brand uses this label -->
  74. ```
  75. ```css
  76. /* deslop-ignore-file: COL-03 the paper colour is part of this brand */
  77. ```
  78. - `deslop-ignore: ID, ID` skips those rules on the line that carries the comment and on the line below it. With no IDs it skips every rule there.
  79. - `deslop-ignore-file: ID, ID` anywhere in a file skips those rules for the whole file. With no IDs it skips the file.
  80. Don't ignore a whole file just to get a clean report.
  81. ## Your own brand
  82. The house style in `SKILL.md` is a default for projects without a brand. To use yours, copy `references/brand.example.md` to `references/brand.md`, fill it in, and put your logo files in `references/brand/`. The skill reads that file first, and it wins wherever it differs from the house style. Both paths are in `.gitignore`.
  83. ## Rules
  84. [`scripts/RULES.md`](scripts/RULES.md) lists every rule with its severity: 81 code rules and 42 screenshot checks. "Certain" rules rarely misfire. "Confirm" rules find a likely tell that has honest uses, so the report lists them separately. The rules live in `scripts/rules.json`; after editing it, run `python3 scripts/make_rules_md.py` to regenerate RULES.md.
  85. ## Tests
  86. ```sh
  87. python3 tests/run_tests.py
  88. ```
  89. The tests run the scanner against the fixtures in `tests/fixtures`, check that every rule fires where it should and stays quiet on look-alikes, test `--url` against a local server, and check that every pattern in the catalogue is covered exactly once.
  90. ## Credits
  91. deslop builds on other people's research and tools: avoid-ai-design by funboy322, unslop by Damir Sokolovsky, Impeccable by Paul Bakaus, Hallmark, design-slop-cop and the research of Adrian Krebs, tasteskill, Uncodixfy, and Anthropic's frontend-design skill. [NOTICE](NOTICE) says what came from each and under which licence. The full list of articles and threads is at the end of `references/patterns.md`.
  92. ## Licence
  93. MIT. Copyright (c) 2026 Pratyush Jain. See [LICENSE](LICENSE). Material adapted from other projects keeps its own licence, as listed in NOTICE: the parts adapted from unslop are under CC BY 4.0.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.