De-Slopping Skills: What They Actually Delete

What it removes for you: turning "this reads like AI" into a deletable list

When you edit an AI draft yourself, most of the time goes into "something's off and I can't say what." This category breaks that sentence into named items. Not "make it flow naturally" — "pattern 8: the dash used as a universal connector, delete it." The four implementations differ in granularity, not in structure.

How to install (commands copied from each project's own page)

# (1) blader/humanizer — the best known one; invoke with /humanizer
npx skills add blader/humanizer --global

# (2) hardikpandya/stop-slop — this is upstream (Skill 16's boraoztunc copy is a mirror)
git clone https://github.com/hardikpandya/stop-slop ~/.claude/skills/stop-slop

# (3) Aboudjem/humanizer-skill — the only one with a recomputable score
npx skills add Aboudjem/humanizer-skill
# Claude Code plugin form: claude plugin install humanizer@10x

# (4) gencharitaci/humanize-skills — install this one for its language catalogs
npx skills add gencharitaci/humanize-skills

The first thing you can actually run

Don't start by asking a model to rewrite anything. Run (3)'s local CLI first — plain Node, zero dependencies, no network calls, and the only thing in this category that produces a result without spending model tokens:

# Score a draft (0–100 AI-tell score, five verdict bands)
node cli/index.js score examples/blog-post/before.md

# Diff before/after and check that numbers, URLs, dates, versions, acronyms survived
node cli/index.js compare --before draft.md --after final.md --check-facts

Run it on something you shipped last week. If the score already looks fine, you can skip this whole category — what you save is the description budget every installed skill charges you on every turn.

Mechanism: why it works, and the opposite answers two authors gave

Three layers

1. A named list demotes taste to lookup. blader/humanizer's 25 patterns come in five groups: A, staging instead of stating ("not X but Y", sayings that sound deep, a staged run-up before the point); B, rhythm by rule (forced triads, repeated sentence openings, dashes as the universal connector); C, inflation and borrowed authority; D, formatting by rule (bold as decoration, decorative headings); E, leftovers from chat and draft (chatbot residue, knowledge-limit disclaimers). The first five are strong enough to justify an edit on a single sighting. The model doesn't have to judge quality — only to find matches.
2. Self-critique is a cheap second pass. The output runs in four steps: mark the tells, draft the rewrite, check that draft, write the final. That third step costs almost nothing and catches a real share of the damage, because rewriting introduces fresh slop of its own.
3. The fidelity layer guards the expensive bug. blader/humanizer's README is explicit: a name, number, date, quote or citation must come from the source or the writer, and if a detail is missing the skill asks rather than inventing one. Aboudjem turned that promise into an executable --check-facts. The worst failure in rewriting isn't a worse draft — it's a better-sounding one. "Up 23.4% year over year" becomes "growth approaching a quarter," which reads better and is no longer the same number.

What makes stop-slop the lightest one

hardikpandya/stop-slop (17.3k stars per its GitHub page, 2026-09; SKILL.md carries v2.0.0 and an internal date of 2026-03-17) splits into SKILL.md plus three references/ files: phrases.md, structures.md, examples.md. Its eight categories are throat-clearing openers, emphasis crutches, business jargon, adverbs, meta-commentary, performative emphasis, telling instead of showing, and vague declaratives. The key detail is that phrases.md is a two-column swap table (navigate → handle, address), which means you can use it without installing anything — paste the table into your own prompt. It also ships a five-dimension score (directness, rhythm, trust, authenticity, density, 1–10 each), with revision warranted below 35/50. That stopping line is exactly what blader's version lacks.

Real use and limits: other languages, and three places to encode your voice

Non-English slop is a different table

"Delete every -ly adverb" and "cut the em dashes" do nothing to a Chinese, Japanese or Turkish draft. gencharitaci/humanize-skills (CC BY 4.0, 0 stars) is the only one here with per-language catalogs: references/languages/ holds 20 files — zh.md, ja.md, ko.md, ar.md, ru.md, tr.md and the rest — plus a modes/ directory with ten purpose modes. The Chinese catalog is concrete enough to use as a checklist straight off: overused vocabulary, reflexive marketing superlatives, sentence-pattern tells (the repeated "not merely X, but rather Y" construction; false-depth reveals of the "at bottom, the real question is…" kind; guided-tour openers like "let us now dive deep into…"), punctuation and number conventions, false-positive carve-outs and one worked before/after example.

Encoding your own voice: three layers, don't pick the wrong one

"Write like me" and "stop sounding like AI" are different jobs. The second runs on lists; the first runs on examples — and which layer you put it in decides when it applies:

• Output style (official; present every turn) — a Markdown file in ~/.claude/output-styles/ or the project's .claude/output-styles/, with name, description and keep-coding-instructions in frontmatter, switched via /output-style <name>. It changes Claude Code's default instructions themselves and ships with every request. Tone and format defaults belong here.
• Skill (loaded on demand) — pulled in only when this particular turn needs a passage cleaned. The humanizers all live in this layer, and they should: you don't want every reply de-slopped.
• CLAUDE.md (project convention) — repository-level house style that travels with the repo, not with you.

The rule in one line: "I re-explain this every turn" goes in an output style; "I need this occasionally" goes in a skill; "this only holds in this repo" goes in CLAUDE.md. Writing your tone default as a skill hands it to a trigger you don't control — and a voice that changes every time is worse than one that's consistently wrong.

Three hard limits

• Don't leave a subtractive skill running over specs, API docs or contracts. In "this parameter applies only in async mode," the "only" is not filler; stop-slop's "cut the adverbs" cuts the precision.
• Anything with numbers in it gets --check-facts after the rewrite, or doesn't get rewritten at all.
• Keep exactly one sibling installed. All four compete for the same trigger surface; which one the model picks is not up to you, and the symptom isn't an error — it's a voice that changes on every run.

Costs and alternatives: if the goal is beating a detector, don't take that fight

Three things you can check

• OpenAI shut down its own detector. The AI Text Classifier launched 2023-01-31 and was withdrawn 2023-07-20 for a low rate of accuracy. On its challenge set it correctly flagged 26% of AI-written text while labeling 9% of human text as AI. The company that builds the models could not build a reliable detector for them — that's the floor of this conversation.
• The false positives land on specific people. Liang and colleagues, in Patterns (2023), ran 7 detectors over 91 TOEFL essays by non-native English writers and measured an average false-positive rate of 61.3%; prompting a model to rewrite the same essays in a native-speaker register dropped it to 11.6%. Read that carefully: a metric a polish pass can move that far was never measuring authorship.
• The conclusions are still turning over. At the 2026 EACL student research workshop, Al Ali, Helcl and Libovický redid the experiment in Czech across three detector families and found no systematic bias against non-native speakers. That doesn't mean detectors are fixed; it means findings in this area flip every couple of years, and your delivery standard shouldn't be attached to them.

What to do instead of installing one

• Move the table into your prompt. stop-slop's references/phrases.md is already a two-column swap table; paste it into an output style or CLAUDE.md and you get zero trigger cost, always on, and no contention with other skills. Best value in this issue.
• Use examples instead of rules. The official brand-guidelines and doc-coauthoring skills hand over sample passages rather than lists of prohibitions — three before/after pairs usually beat ten rules. Skill 18 is entirely about that recipe.
• Three rules of your own will do. Read your last five pieces, find the three constructions that annoy you most, write them down. They beat any 55-item list, because those three are the ones you actually commit.