Projects / ne-slop-text
ne-slop-text
A Claude Code skill for writing text that reads as written by a particular person.
- three models writing the same page with no method
- first draft
- after the revision
- Released
- 7 November 2026
- Runs in
- Claude Code. The scripts need Python 3 and nothing else.
- Source
- GitHub
- Cost
- 6 to 8 agent calls and a few minutes for a longer text
git clone https://github.com/boris-timofeenko/ne-slop-text ~/.claude/skills/ne-slop-textThen ask Claude Code to write “ne-slop”, or run /ne-slop-text.
I made a quest game, Belgrade 2026, and its graphics came out alive. The hero was cut out of a friend’s photo with the white fringe of the magic eraser left on. That became my ne-slop skill for creative work. For text it was not enough, and text run through the popular humanizer skill I could spot from a mile away.
I built ne-slop-text with Claude on my own texts and tested it blind on myself. For each task three fresh agents on different models write the text with no method, and an annotator marks the draft and these defaults on a questionnaire of 26 discourse features, seeded from the StoryScope paper (Russell et al., COLM 2026), where narrative decisions alone told AI from human stories at 93% F1.
In the last round, a message to a hiring team, four default writers got the same material, and the two Opus letters had the same skeleton down to the closing line, “I’d be glad to have that coffee”. Where the draft did the same machine-like things as every default, the reviser cut three clever turns and three of four phrases quoted from the job ad. The draft moved from third place to first. The plain Opus letter came last, and I recognised it.
That is one judge, three tasks and one domain so far, so the numbers show a direction.
Screenshot
