anti-slop

Eliminate AI-sounding patterns from any written output. Applies editorial rules synthesized from the best open-source anti-slop tools: banned phrases, structural pattern detection, false agency checks

SKILL.md

GitHub

Anti-Slop Writing Quality Gate

Eliminate recognizable AI patterns from text. Synthesized from the best open-source anti-slop tools.

Core Philosophy

LLMs collapse toward defaults — the same phrases, structures, and rhythms. There are two fixes:

  1. Remove known bad patterns (this skill)
  2. Add enough context that the model makes informed, original choices

Telling a model to "write better" just creates new slop. Listing what NOT to do forces genuine novelty.

When to Apply

Run this check on ANY text before it ships — blog posts, tweets, emails, docs, marketing copy. Load references/patterns.md for the full ruleset.

Quick Audit (5-Point Check)

Before publishing, verify:

  1. No banned phrases — scan for the ~30 phrases listed in patterns.md
  2. No false agency — inanimate things don't "tell us", "reward", or "demand"
  3. No em-dash abuse — zero em-dashes (use commas, periods, or parentheses instead)
  4. Specific over vague — numbers, names, dates instead of "various", "numerous", "significant"
  5. Voice present — does it sound like a specific person wrote it, or could any AI have produced it?

Scoring Rubric (from stop-slop)

Score each piece out of 50. Below 35 = mandatory revision.

Category Points What to Check
Authenticity /10 Genuine voice? Opinions? First-person where appropriate?
Specificity /10 Concrete details, numbers, names? Or vague generalities?
Structure /10 Varied sentence length? No formulaic patterns?
Word choice /10 Fresh vocabulary? No crutch words?
Flow /10 Natural rhythm? Reads like speech, not a textbook?

Two-Pass Revision Process (from humanizer)

Pass 1 — Rewrite: Apply all rules from patterns.md. Remove banned phrases, fix false agency, vary rhythm.

Pass 2 — Self-critique: Ask "What still makes this obviously AI-generated?" about your own output. Fix what you find. The self-critique loop catches patterns that survive the first edit.

References

  • patterns.md — Complete banned phrase list, structural patterns to avoid, and detection rules. Load when auditing text.

Attribution

This skill synthesizes rules from four open-source tools. All credit to the original authors:

  • humanizer by blader (MIT, 10,300+ stars) — two-pass audit, 25 pattern categories
  • stop-slop by Hardik Pandya (MIT, 2,100+ stars) — banned phrases, scoring rubric
  • unslop by Matt Shumer (MIT, 180+ stars) — empirical default detection
  • impeccable by Paul Bakaus (Apache 2.0, 11,700+ stars) — design slop detection

Curated by shannhk.