{"id":"ae93e6cf-b993-44a1-a64a-f6846e1f2f3b","slug":"llmstrip","kind":"skill","name":"llmstrip","summary":"Strip LLM writing patterns from prose, code comments, and commits — CLI, Claude Code skill, Cursor rule, and LLM system prompt.","owner":"HugoLopes45","repoUrl":"https://github.com/HugoLopes45/llmstrip","license":"MIT","licenseTier":"permissive","trust":"community","stars":12,"match":null,"fewStars":false,"format":"skill","createdAt":"2026-02-20T16:57:44.000Z","addedAt":"2026-10-14T05:56:26.114Z","logoUrl":"https://avatars.githubusercontent.com/u/4919100?u=116ec755980f5deca9d1c3ed2e005e661624587c&v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["developer-tools","writing"],"tags":["ai-tools","ai-writing","anthropic","claude","claude-code","cli","cursor","developer-tools"],"runtime":"binary","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-17T09:22:40.443Z","sample":false,"url":"https://conduits.dev/tools/llmstrip","language":"Rust","transports":["stdio"],"install":{},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"# llmstrip\n\nStrip AI patterns from text and code. Prompt + linter.\n\n> 25× excess frequency of `delve` post-ChatGPT. 280+ excess words identified. 34 structural rules. Grounded in peer-reviewed data.\n\n![demo](assets/demo.gif)\n\n## Two ways to use it\n\n**Interactive — the prompt (recommended)**\n\nPaste the system prompt into any LLM, or use the Claude Code skill. The LLM rewrites your text using all 34 rules, including structural patterns that no regex can catch.\n\n**Automation — the binary**\n\nRun `llmstrip --report --fail` in CI or as a git hook. Same rules, deterministic, no network, no runtime.\n\n## Install the prompt\n\n**Claude Code skill:**\n\n```bash\nmkdir -p ~/.claude/skills/llmstrip\ncurl -sL https://raw.githubusercontent.com/HugoLopes45/llmstrip/main/prompts/claude-code.md \\\n  > ~/.claude/skills/llmstrip/SKILL.md\n```\n\nOr from the repo: `make install-skill`\n\nType `/llmstrip` in any Claude Code session to clean the current file.\n\n**Other tools:**\n\n```bash\nmake install-cursor    # Cursor — .cursor/rules/llmstrip.mdc\nmake install-copilot   # GitHub Copilot — .github/copilot-instructions.md\nmake install-windsurf  # Windsurf\nmake install-zed       # Zed\nmake install-all       # All of the above\n```\n\nOr copy `prompts/system-prompt.md` into any LLM's system prompt directly.\n\n## Install the binary\n\n```bash\ncurl -fsSL https://raw.githubusercontent.com/HugoLopes45/llmstrip/main/scripts/install.sh | sh\n```\n\nOr via Cargo:\n\n```bash\ncargo install --git https://github.com/HugoLopes45/llmstrip\n```\n\n## Use cases\n\n### Filter every Claude response automatically\n\n```bash\n# .claude/settings.json — PostToolUse hook\n{\n  \"hooks\": {\n    \"PostToolUse\": [{\n      \"matcher\": \".*\",\n      \"hooks\": [{\"type\": \"command\", \"command\": \"llmstrip\"}]\n    }]\n  }\n}\n```\n\n### Strip AI prose\n\n```bash\necho \"Let me delve into this robust and comprehensive approach.\" | llmstrip\n# -> Let me dig into this solid and thorough approach.\n```\n\n### Clean AI-generated code comments\n\n```bash\nllmstrip --report --mode code service.py\n\n# Mode: code  |  4 finding(s)\n#\n# HIGH (1)\n#   line 2: LLM docstring boilerplate: 'this function serves as'\n#\n# MEDIUM (3)\n#   line 1: Type-in-name: use 'user' instead of 'userDataObject'\n```\n\n### Catch AI commit messages before they land\n\n```bash\n# .git/hooks/commit-msg\n#!/bin/sh\nllmstrip --mode code --rules commits --report --fail --min-severity high \"$1\"\n```\n\n## CI gate\n\nBlock AI-written release notes in CI:\n\n```bash\nllmstrip --report --fail --min-severity high release-notes.md\n```\n\n## Rules\n\n34 rules across two levels:\n\n**Word-level (Rules 1-24):** significance inflation, banned vocabulary (delve, robust, leveraging, seamlessly...), copula avoidance, sycophantic openers, chatbot closers, filler phrases, excessive hedging.\n\n**Structural (Rules 25-34):** compound clause addiction, mini-essay paragraphs, sentence-initial transition saturation, paired construction overuse, noun phrase bloat, self-congratulation framing, abstraction level monotony, missing contractions, forward projection cliche, resume verbs.\n\nFull list: [`rules/`](rules/)\n\n## Research\n\nRules are grounded in peer-reviewed corpus studies:\n\n- **Kobak et al. (2025)** — 15M PubMed abstracts. `delve` appeared 25× more often post-ChatGPT. 280+ excess words. [arXiv:2406.07016](https://arxiv.org/abs/2406.07016)\n- **Liang et al. (2024)** — 950K+ papers. `pivotal`, `intricate`, `realm` doubled post-2023. [arXiv:2404.01268](https://arxiv.org/abs/2404.01268)\n- **Juzek & Ward (2025)** — RLHF causes it. Human raters prefer formal-sounding output. Models overfit. [arXiv:2412.11385](https://arxiv.org/abs/2412.11385)\n\n## Contribute a pattern\n\nSee [CONTRIBUTING.md](CONTRIBUTING.md). Open an issue with the pattern, a before/after, and a corpus source if you have one. Label it `new-rule`.\n\n## License\n\nMIT.\n"}