This repo holds three Reddit-mined studies of what people flag as AI "slop", and a Claude skill built from each: one for websites and UI, one for writing, and one for code. The flagship study, written up in full below, ranked the visual "tells" people use to spot AI-built (vibe-coded) websites. It mines public Reddit discussion from the free Arctic Shift archive, tabulates which design features get named most, and verifies the findings against real quotes.

Everything here is reproducible with Python and the standard library plus matplotlib. No API key, no auth.

Pick the one for what you want to clean up. Each is a Claude skill built from its study: a SKILL.md with build and audit modes, a tells catalog ranked by this repo's data, and a standalone scanner that flags the tells and gates CI on its exit code.

• unslop-ui removes the cues that make a website read as AI-generated: the default shadcn/Tailwind look, AI-purple gradients, gradient hero text, unprompted neon glow, emoji-as-icons, and the centered-hero-plus-three-cards layout. Install: unzip skill/unslop-ui.skill -d ~/.claude/skills/, or upload skill/unslop-ui.skill in the claude.ai skills UI.
• unslop-text removes the cues that make prose read as AI-written: the em dash, the "it's not just X, it's Y" cadence, leftover assistant boilerplate, sycophantic openers, the delve/leverage diction, and the "in conclusion" wrap-up. Install: unzip unslop-ai-text/skill/unslop-text.skill -d ~/.claude/skills/, or upload unslop-ai-text/skill/unslop-text.skill in the claude.ai skills UI.
• unslop-code removes the tells that make source code read as AI-written (leftover chat artifacts, placeholder comments, emoji, swallowed errors, narrating comments, generic placeholder names) and points you at the structural tells a linter passes, like boilerplate and hallucinated APIs. Install: unzip unslop-ai-code/skill/unslop-code.skill -d ~/.claude/skills/, or upload unslop-ai-code/skill/unslop-code.skill in the claude.ai skills UI.

• 3,214,533 posts scanned across 47 AI and SaaS subreddits, 2020 to 2026.
• 46,971 of those are on-topic (about AI-built sites), 1.46% of the scanned base.
• 3,033 comments harvested from 125 canonical threads ("why do AI sites all look the same", "dead giveaways for AI slop websites", and similar). These comments are the cleanest signal because they are 100% on-topic.
• Every top tell was adversarially verified by an independent pass. 11 of 12 held up; one ("mesh / blob / aurora backgrounds") was rejected as a keyword artifact.

The loudest complaint is not any single feature. It is that the sites are recognizable on sight. "They all look the same" and "screams AI / slop" each show up in about 13% of on-topic posts. Among specific features, the ranking by share of on-topic comments is:

shadcn/Tailwind defaults and the "AI purple" gradient lead. The stereotypical Twitter memes (bento grids, glassmorphism, aurora gradients) sit near the bottom or get rejected.

The conversation barely existed before 2024. Measured as share of posts (not raw counts, which just track subreddit growth), it jumped roughly 150x from 2023 to 2024.

The same findings are packaged as unslop-ui, a Claude skill that strips these patterns while building or auditing a site. It does not impose a look. It removes the AI tells (including the newer cream-plus-serif-plus-sage "tasteful default" that just trading one default for another produces) and forces a deliberate, project-specific choice instead. It includes a standalone scanner (skill/scripts/devibe_scan.py) that greps a codebase, prints findings with a vibe score, and gates CI on the exit code. See skill/README.md to install it or run the scanner, and the animated demo, where one prompt becomes four distinct deliberate designs that all pass the scanner.

Run in this order. Each script is sequential and resumable (the harvesters checkpoint and dedupe by id), and writes its outputs into this folder.

pip install -r requirements.txt
cd unslop-ai-ui

python3 collect.py            # Phase 1-2: per-sub totals + matched-by-year (aggregate endpoint)
python3 harvest.py 3000       # Phase 3: harvest on-topic post text -> corpus.jsonl
python3 harvest_comments.py   # comments from the canonical threads -> comments.jsonl
python3 analyze.py            # post-level tell tabulation + the first five charts
python3 analyze_comments.py   # comment-level tell tabulation (the cleaner ranking)
python3 make_charts.py        # honest comment-level + comparison charts
python3 make_charts2.py       # scale, raw counts, funnel, concentration, co-occurrence, sentiment, threads

The committed corpus.jsonl.gz is a snapshot. To run the analysis scripts against it without re-harvesting, gunzip corpus.jsonl.gz first. post_workflow.js is the multi-agent verification and drafting workflow used to vet the tells.

The study's data, scripts, and charts live in unslop-ai-ui/:

• Scripts: collect.py, harvest.py, harvest_comments.py, analyze.py, analyze_comments.py, make_charts.py, make_charts2.py, post_workflow.js.
• Raw data: corpus.jsonl.gz (46,971 posts), comments.jsonl (3,033 comments). Fields: id, subreddit, created_utc, score, title/selftext or body, permalink. No usernames were collected.
• Tables: comment_tell_counts.csv and tell_counts.csv (the rankings), scanned_totals_by_sub.csv, totals_by_year.csv, matched_by_year.csv, growth_by_year.csv, tell_share_by_year.csv, harvest_ledger.csv, summary.txt.
• Quote banks: tell_examples.md and comment_tell_examples.md (verbatim quotes with permalinks).
• Charts: twelve PNGs.
• DATA_AND_GRAPHS.md: the full master table, growth table, and chart index.

Share of posts or comments, not raw counts. A tell that recurs across many threads ranks above one that spikes in a single viral thread. Tells are detected with a synonym lexicon (see unslop-ai-ui/analyze.py), counted over a design-context subset of the corpus, and the comment-level numbers are treated as the primary ranking because those threads are all on-topic.

This is a proxy for vocal, online opinion, so trust the relative ordering more than the exact percentages. Small subreddits are noisy, and keyword matching can miss sarcasm or catch the wrong sense of a word. See unslop-ai-ui/DATA_AND_GRAPHS.md for the per-tell false-positive notes.

Code is MIT (see LICENSE). The harvested text is public Reddit content collected via Arctic Shift and belongs to its original authors; see unslop-ai-ui/DATA_NOTE.md.

The same Reddit-mining method has been applied to two other media. Each lives in its own folder with its scripts, verified findings, quote bank, and a skill/ folder scaffolded for an accompanying Claude skill (packaged separately):

• unslop-ai-code/ — the tells that give away AI-written code, ranked and adversarially verified the same way. Ships its corpus as corpus.jsonl.gz.
• unslop-ai-text/ — the tells that give away AI-written text (prose, marketing, academic). Its raw corpus is too large to ship, so it regenerates from collect.py (see that folder's DATA_NOTE.md).