detect-ai-slop

Use when the user wants to SEE which AI tells are in a piece of text without rewriting it. Triggers include "flag the AI slop", "where are the tells", "show me what to fix", "detect AI-generated writi

SKILL.md

GitHub

Detect AI Slop

Report where the AI tells are. Return located findings, grouped by category, so the author knows exactly what to fix. This lens never rewrites and never emits a single headline score — a number invites gaming, a located finding invites editing.

Rules, categories, severity, and the override-merge order are defined once in ../../shared/spec.md. Read it before reporting. The countable tells are scored by a deterministic script; the meaning-dependent ones are yours to judge.

Two tiers

Deterministic tier (preferred). Run the bundled script. It counts the countable tells (vocabulary-table hits, em-dash density, fixed templates, bold/bullet soup, fixed scaffolding phrases) and returns stable JSON.

python3 shared/slop_count.py --file <path> \
  [--allow-list ~/.claude/config/remove-ai-slop/user-allow-list.md] \
  [--allow-list .claude/skills/remove-ai-slop/overrides/user-allow-list.md] \
  [--extra-vocab ~/.claude/config/remove-ai-slop/user-vocabulary.md] \
  [--extra-vocab .claude/skills/remove-ai-slop/overrides/user-vocabulary.md]

Pipe text on stdin instead of --file when the user pasted it. Pass whichever override files exist (skip silently if absent). The script is the sole parser of the override files: consume its findings and summary, do not re-parse the override tables yourself.

Judged tier (you). The script cannot see meaning. After the script runs, add findings for the meaning-dependent categories from the spec — padding-and-filler, elegant-variation, and contextual cases (is robust actually wrong here? is this copula avoidance or a legitimate verb?). Mark these tier: judged.

Fallback — when python3 is absent

A Claude Code plugin has no guaranteed runtime. If python3 is not available (the run fails or command -v python3 is empty), do the whole detection yourself from the spec and references, read the allow-list files directly (the same files the script would have read), and tag the result unstable — say plainly: "No python3 found; this is an LLM-only pass and the counts are not byte-stable." Do not run two parsers at once.

Output

Group findings by category, most severe first (severity is in the spec). For each:

[high] formulaic-constructions — para 3
  "It's not just a tool, it's a way of life."
[low] overused-vocabulary — para 1
  "delve", "leverage", "tapestry"

Rules:

  • No single 0-10 number. A per-category tally (from the script summary) is fine.
  • Quote the offending text and give the paragraph index. That is the anchor.
  • If the script ran, say so and note the result is stable. On fallback, tag it unstable.
  • If there are no findings, say: *No AI tells detected.* Do not invent problems.
  • Respect the allow-list. A word the user allow-listed is never a finding.

Scope

Report only. To rewrite the text, that is remove-ai-slop. To get a per-category breakdown framed as a quality pass, that is score-ai-slop (ships next). Do not expand beyond reporting what is there.