ai-slop-detector
Installation
SKILL.md
This skill is the audit procedure for the norm output_discipline states in ai-prompts/CLAUDE.md: that file says
what to avoid, this one how to find an instance already in prose or code. Read that contract first. Judge words
in context; the patterns below identify candidates, not banned tokens.
Prose tells: what greps and what doesn't
Scan lexically, then judge each hit in context before reporting it. Run the scan over the target file, not a diff, since slop introduced gradually never shows in any single diff. Zero hits is not evidence the text carries content: prose with every listed token removed can still state nothing, so the judgment pass runs regardless.
| Tell | Pattern (case-insensitive) |
|---|---|
| Announcement / closing restatement | in this (section|article|document)|^overall,|in summary|it is worth noting |
| Empty intensifier / self-praise | \b(robust|comprehensive|seamless|successfully|significantly)\b |
| Informationless hedge | \b(essentially|basically|arguably)\b |
| Formulaic parallelism | not only .* but also|it('s| is) not just .*, it('s| is) |
| Sycophantic opener | ^(you('re| are) absolutely right|great question|excellent point) |
| Em dash (English prose) | the em dash character, U+2014: perl -CSD -ne 'print "$.: $_" if /\x{2014}/' |
| Decorative emoji | grep -P '[\x{1F300}-\x{1FAFF}\x{2600}-\x{2604}\x{2606}-\x{27BF}]' |