Built an AI slop detector using your design-checklist and AI slop blacklist

[dakevinci]

Hey Garry — built vibecheck-slop-stopper, a tool that detects AI-generated code in PRs (GitHub Action + CLI + Claude Code skill).

Your gstack design-checklist and AI Slop Blacklist were a primary source for our visual slop detection (P11) and AI language detection (P9):

From your AI Slop Blacklist (10 items) → our grep rules:

• Purple/violet/indigo gradient detection → G120, G121
• Centered text density >60% → G122
• Uniform bubbly border-radius → G123
• Generic hero copy ("Unlock the power of...") → G124
• Emoji as design elements → G125
• Colored left-border cards → G126

From your banned vocabulary list (30+ words) → G110:

• Co-occurrence model: flag when 3+ banned LLM words appear in the same file's comments, not individual matches

From your design scoring framework:

• Confidence tiers (HIGH/MEDIUM/LOW) mapped to our FP risk ratings

We turned these into a TOML atom/stanza engine — human-readable pattern definitions that compile to ripgrep regex at scan time. 72 rules across 11 categories, runs as a GitHub Action on every PR.

You're credited in our README. MIT licensed.

Thanks for open-sourcing gstack — the AI slop blacklist in particular is the best codified list of AI visual tells I've seen.

[raye-deng]

Nice work building on the gstack patterns. Your co-occurrence model (flag when 3+ banned LLM words appear together) is a smart improvement over per-word matching — it dramatically reduces false positives while catching the distinctive AI-generated prose style.

One category your 72 rules might not cover yet: code-level AI artifacts. The gstack checklist focuses on visual/UI slop, but AI-generated code has its own tells that are detectable statically:

• Hallucinated imports — packages that don't exist on npm/PyPI but look plausible (e.g., @radix-ui/react-dropdown instead of @radix-ui/react-dropdown-menu)
• Deprecated API calls — AI models trained on older docs using removed methods (e.g., jest.mock() patterns from v27, vitest APIs removed in v4)
• Over-abstracted utility extraction — creating helpers/ or utils/ files for functions called exactly once

These pass every linter and type checker but break at runtime. We built a scanner focused on these code-level patterns:

npx open-code-review@latest ./src

It validates imports against real registries and checks for deprecated API usage with version awareness. Could be complementary to your grep-based visual slop detection — you catch the UI/code style issues, it catches the dependency and API correctness issues.

The TOML atom/stanza engine approach you took is elegant. Have you considered making it extensible so others can contribute rule sets? A community-maintained rule catalog could evolve faster than any single team's pattern library.