llmstrip

Strip AI patterns from text and code. Prompt + linter.

25× excess frequency of delve post-ChatGPT. 280+ excess words identified. 34 structural rules. Grounded in peer-reviewed data.

Two ways to use it

Interactive — the prompt (recommended)

Paste 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.

Automation — the binary

Run llmstrip --report --fail in CI or as a git hook. Same rules, deterministic, no network, no runtime.

Install the prompt

Claude Code skill:

mkdir -p ~/.claude/skills/llmstrip
curl -sL https://raw.githubusercontent.com/HugoLopes45/llmstrip/main/prompts/claude-code.md \
  > ~/.claude/skills/llmstrip/SKILL.md

Or from the repo: make install-skill

Type /llmstrip in any Claude Code session to clean the current file.

Other tools:

make install-cursor    # Cursor — .cursor/rules/llmstrip.mdc
make install-copilot   # GitHub Copilot — .github/copilot-instructions.md
make install-windsurf  # Windsurf
make install-zed       # Zed
make install-all       # All of the above

Or copy prompts/system-prompt.md into any LLM's system prompt directly.

Install the binary

curl -fsSL https://raw.githubusercontent.com/HugoLopes45/llmstrip/main/scripts/install.sh | sh

Or via Cargo:

cargo install --git https://github.com/HugoLopes45/llmstrip

Use cases

Filter every Claude response automatically

# .claude/settings.json — PostToolUse hook
{
  "hooks": {
    "PostToolUse": [{
      "matcher": ".*",
      "hooks": [{"type": "command", "command": "llmstrip"}]
    }]
  }
}

Strip AI prose

echo "Let me delve into this robust and comprehensive approach." | llmstrip
# -> Let me dig into this solid and thorough approach.

Clean AI-generated code comments

llmstrip --report --mode code service.py

# Mode: code  |  4 finding(s)
#
# HIGH (1)
#   line 2: LLM docstring boilerplate: 'this function serves as'
#
# MEDIUM (3)
#   line 1: Type-in-name: use 'user' instead of 'userDataObject'

Catch AI commit messages before they land

# .git/hooks/commit-msg
#!/bin/sh
llmstrip --mode code --rules commits --report --fail --min-severity high "$1"

CI gate

Block AI-written release notes in CI:

llmstrip --report --fail --min-severity high release-notes.md

Rules

34 rules across two levels:

Word-level (Rules 1-24): significance inflation, banned vocabulary (delve, robust, leveraging, seamlessly...), copula avoidance, sycophantic openers, chatbot closers, filler phrases, excessive hedging.

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.

Full list: rules/

Research

Rules are grounded in peer-reviewed corpus studies:

• Kobak et al. (2025) — 15M PubMed abstracts. delve appeared 25× more often post-ChatGPT. 280+ excess words. arXiv:2406.07016
• Liang et al. (2024) — 950K+ papers. pivotal, intricate, realm doubled post-2023. arXiv:2404.01268
• Juzek & Ward (2025) — RLHF causes it. Human raters prefer formal-sounding output. Models overfit. arXiv:2412.11385

Contribute a pattern

See CONTRIBUTING.md. Open an issue with the pattern, a before/after, and a corpus source if you have one. Label it new-rule.

License

MIT.