Deslop Claude skill released to strip LLM prose tells
Start here
A developer has released Deslop, a Claude skill that revises drafts to remove the fingerprints of LLM-generated prose. Its rules are drawn from corpus-linguistics and psycholinguistics research and checked against three public corpora in which humans and models answered the same prompts. In the author's example, a 228-word client email containing 3 Claude tells became 132 words with 0 tells and no facts dropped after one pass. Users install it by copying the folder into ~/.claude/skills/deslop, then asking Claude to deslop a draft. The core scanner scripts need only Python 3; optional extras such as deep_scan.py and judge_prompt.py require a Python 3.12 virtual environment with spaCy, torch, transformers and sentence-transformers. The bundled third-party text under sources/raw/ and sources/papers/ is for local analysis only and is not to be redistributed.
Generated from reports · updated 2 h ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIDeslop: a Claude skill that strips LLM tells from drafts using corpus and psycholinguistic research
Deslop is a Claude skill that revises drafts to remove the fingerprints of LLM-generated prose, with rules drawn from corpus-linguistics and psycholinguistics research and checked against three public corpora where humans and models answered the same prompts. In the example, a 228-word client email with 3 Claude tells became 132 words with 0 tells and no facts dropped after one pass. Install by copying the folder into ~/.claude/skills/deslop, then ask Claude to deslop a draft; the core scanner scripts need only Python 3, while optional extras such as deep_scan.py and judge_prompt.py require a Python 3.12 venv with spaCy, torch, transformers and sentence-transformers.
Heat for this event
Not enough continuous observations yet to draw a trend.