Best AI writing cleanup skills for AI agents

Ask an agent to "make this sound human" and it has no definition of the job. It will cut a few hedges, swap a few words, and hand back prose that reads about the same, because the model grading the draft is the model that wrote it. That is the actual failure: the draft contains em dashes the model does not remember writing, and "sounds better now" is not a check anyone can run. Worse, the rewrite quietly reintroduces what it just removed — take out "it's not just X, it's Y" and you get "this is less about X than Y" back in its place.

The split that matters across these documents is not how strict they are. It is whether the document tells you what it will change in your text, and whether it says what it is for. A few name thresholds you can count and boundaries they refuse to cross. Several are a mood. And a file called humanize may not be an editor at all: one of them ships your draft to a commercial endpoint and meters it at a credit per word.

Top picks

Start with the rulebook that hands your agent countable rules

harshaneel/humanize/humanize opens with seven hard rules and then tells the agent to enforce them at the very end, against the finished draft, by counting. Em dashes: at most one per 300 words, and zero under 300 words. Sentence-length spread: the longest sentence must beat the shortest by 20 or more words, and fewer than half the sentences may sit in the 10-to-20-word band. The pre-output gate makes the agent write each count down, zeros included, on the reasoning that an unenumerated "looks clean" always passes. Later it asks for the word count of every sentence in order, as a list, because a read-through always sounds varied to the model that produced it. It also states what it will not do: guarantee a detector score, add invented facts for specificity, or change the meaning of the input. Be clear about what it says it is for. Its description names both "make this sound more human" and rewriting to avoid AI detection, and the table its rules rest on sits under a section called "Mental model: what detectors actually measure" — perplexity, burstiness, hedge density, punctuation fingerprint. The rules themselves are ordinary style rules with numbers attached to them. Two of the optional advanced techniques at the end are the exception, since both put a detector in the loop and so need tooling most readers will not have. That same section closes on a short list of tricks the document tells the agent not to bother with, homoglyph injection and watermark stripping among them. Naming those in order to rule them out is part of why this one reads as honest. The rival documents here are genuinely useful teaching material — softaworks/agent-toolkit/humanizer and pproenca/dot-skills/humanize pair a named pattern with a real before-and-after rewrite, and poteto/noodle/unslop compresses the same catalogue into a page you can read in one sitting — but "em dash overuse" in the softaworks and pproenca documents carries no threshold, so the agent decides what overuse means, and unslop settles it by banning the character outright, which is an answer rather than a measurement. Caveat: this one is long, and its front matter declares no license (the repository record says MIT).

For anything going out the door, use the pass that re-scans itself

JuliusBrussee/skills/fuck-slop is built around the observation that the worst tells survive a rewrite, so it makes the scan mechanical and then runs it again on the output. Negative parallelism gets a three-way triage: if nobody actually believes the X half, delete it and assert Y with whatever evidence exists; if the contrast is real, name who holds it; if the sentence asserts nothing, delete the sentence, which the document says is most cases. It then lists the escape hatches that count as fresh findings — "less about X than Y", "the real X is Y", "the question isn't X, it's Y" — so the agent cannot launder the pattern into a synonym. The verify pass is capped at four rounds, after which the sentence gets rebuilt from its bare claim. Two rules make it safe to point at a real draft: never invent specifics (leave a marked placeholder instead), and no fake typos or forced slang, because humanizer-tool output is its own genre of slop. alecs5am/ralphy/humanizer reaches the same place by the opposite route, and its worked example is the clearest demonstration in this family of why the rule exists. Its draft rewrite answers a puffy input with a named study, a 55% figure and two named interviewees, none of them anywhere in the input. Then the document's own audit step flags exactly that, warning that invented-looking names and citations are a liability unless somebody can actually source them, and the final rewrite drops every one of them. Both documents refuse the invention; the difference is where the guard sits. fuck-slop makes it a standing rule the rewrite obeys, with a marked placeholder to leave behind. alecs5am makes it something the agent has to catch on review, which works in a scripted example and is one more step to skip on a real draft. Caveat: fuck-slop's pattern catalogue and genre profiles live in separate reference files, so a partial copy loses the regexes.

When it has to sound like one person, not "human in general"

guia-matthieu/clawfu-skills/voice-injection-rewriter refuses to start without a voice to inject, and grades its own inputs: a stored brand profile is best, an analysis over ten or more samples is very good, three to five pasted samples are good, and a one-line description of your register is rated minimum viable. The rewrite runs five passes — kill the default vocabulary, break rhythm uniformity, inject the signature moves from the profile, match contraction level and pronoun, then read the whole thing aloud asking whether this specific person would actually say each sentence. Output is a change log you can audit: what was swapped, what was removed, what voice marker went in, and a consistency rating. It also prints a list of what it cannot do, including creating a voice from nothing and guaranteeing undetectable output. Branded-Mayhem-Collective-LLC/humanize-kit/humanize also loads a saved voice profile, but its frame is different. It scores the draft 0 to 3 on each row of its own 19-row table of tells, totals the result out of 57, and picks how aggressive the rewrite should be from that number. Nothing is sent anywhere: the table is the document's own account of the patterns it attributes to Pangram, GPTZero and Copyleaks, and its stated purpose is breaking them. It is honest about the limit — its own description says pasting the output into a classifier does not reliably pass — but it also tells the agent never to mention that the text was processed, which is a different instruction than anything in the picks above. It is also the softer prerequisite of the two: it asks for a voice profile up front and then says plainly that it will run without one and hand back something generic. Caveat: the two voice sources voice-injection-rewriter rates highest both need ClawFu tooling this document does not contain — a pre-stored profile fetched with get_brand, or the output of its companion brand-voice-learner. The sources you can supply yourself, pasted samples or a one-line description, are the ones it rates lower.

For judging someone else's text, use the one built against false accusations

mike-coulbourn/claude-vibes/ai-writing-detection is the skill to hand an agent when the question is whether a document was machine-written, and its most useful pages are the ones about not being wrong. It separates definitive artifacts from style: turn0search and oaicite reference markers, utm_source=chatgpt.com inside a link, grok_card tags — copy-paste residue that means something — from vocabulary and rhythm, which mean much less. Then it puts guardrails on the verdict: a 200-word minimum, no flagging on a single indicator, at least three corroborating signals, and a named list of indicators to stop using, including perfect grammar, "bland" prose, and unusual vocabulary. It names the groups most likely to be falsely accused, non-native English speakers among them, and states plainly that false accusations cause real harm. The contrast with humanizerai/agent-skills/detect-ai is stark: that one POSTs your text to a vendor API and prints the API's overall score out of 100, its verdict field, and a five-row table telling you what each score range means, with no guidance on what the number cannot support. Caveats: the 61% false-positive figure it cites for non-native speakers is attributed at the point of use only to "research" — the document does list its sources at the end, Stanford HAI among them, so the reference is chaseable, but chase it before you repeat the figure anywhere it matters; most of its word lists sit in eight reference files alongside the skill; and it has not been updated since 2025-12-21.

For a paper, use the one that states who it is for

momo2young/humanize-academic-writing/humanize-academic-writing puts an academic integrity statement above its instructions, and it is specific rather than decorative: revising your own AI-drafted text and helping non-native English speakers are marked acceptable; using AI to generate ideas you do not understand, and submitting work that does not represent your own intellectual contribution, are marked not. It then does something the general-purpose documents do not attempt: it ships analysis scripts, one for pattern detection with per-paragraph output and a JSON mode, one for before-and-after metrics — type-token ratio, transition density, passive-voice percentage, sentence-length variance — and it requires a one-or-two-sentence rationale for each rewritten paragraph, so the writer learns the pattern rather than outsourcing it. Its don't-list is equally concrete, down to refusing to introduce deliberate grammatical errors. That last rule is a real fork in this space: the humanize-kit document above treats clean grammar as a tell and recommends deliberate fragments in response, while voice-injection-rewriter rejects manufactured imperfection on the grounds that it is an arms race you lose. Register is the reason to choose it: harshaneel's domain calibration covers technical, narrative, creative, business and Slack writing, and unslop covers none, so neither knows that hedging is correct in a findings section. Caveat: the scripts and extended examples live in the repository, and the workflow assumes you can run Python.

Two skills with the same name are not two versions of the same skill

Skill names are community-chosen and not namespaced, so a name tells you the topic and nothing else. Two failure modes follow, and both are easy to check.

The first is the same document wearing two names. The softaworks and pproenca documents above are the same document: the same patterns in the same order, the same worked examples, down to the same rewrite of a sentence about the Statistical Institute of Catalonia. Three lines separate them — the skill's name, a parenthesis in the description, and a credit line to @blader that one front matter carries and the other does not. Choosing between those on publisher reputation is choosing between a file and itself.

The second is the same name over a different job. One skill called humanize edits the text in front of it using rules you can read. Another, humanizerai/agent-skills/humanize, sends your draft to a remote API, needs an API key in your environment, charges one credit per word, and offers an intensity setting its own table labels Bypass. Its neighbours under the same account include a readability calculator and a word counter, published in the same style. The account slug is not a claim about who wrote anything, and the file name is not a claim about what runs.

The check is the same in both cases: open the body before you install. What does this document change in the text, and where does the text go?

Read for three things before you install

Does it commit to a number? A per-300-word em-dash budget and a required sentence-length spread are instructions an agent can verify against its own output. "Avoid em dash overuse" is an opinion the agent has to invent a threshold for, and it will invent a lenient one.

Does it agree with itself, and with its own record? brandonwise/humanizer/humanizer advertises 28 pattern detectors and 560+ vocabulary terms in its description, while its closing process step tells the agent to run detection over 24 patterns and 500+ terms. alecs5am/ralphy/humanizer declares license: MIT in its own front matter, while the repository record for it says Apache-2.0 — for anything you plan to ship, that disagreement is the one to resolve before you install. And two documents still carry the checklist their author wrote them against: EliasOulkadi/shokunin/humanize closes with items like "YAML frontmatter is valid", and besoeasy/open-skills/humanizer opens with a note that the skill name should match its folder. Neither is a reason to reject the file; both tell you how closely it was read before publishing.

Does the runnable part agree with the prose? besoeasy also ships a runnable sed script, offered as a starter for rewriting a whole file at once, that turns every em dash into a comma and every "serves as" into "is". Its own Troubleshooting section says the opposite: "Run pattern scan first, then rewrite only flagged spans; avoid global synonym swaps." Its validation workflow then asks you to "Compare original vs rewritten sentence-by-sentence for factual equivalence." The prose knows a find-and-replace cannot tell which em dash was doing work. The script in the same file does it anyway, and the script is the part that runs.

What you have now

The agent that was going to "make it sound human" now has a definition of the job. Install harshaneel/humanize/humanize as the default rulebook, because it hands the agent counts instead of adjectives and tells it what the rewrite is not allowed to do. Add fuck-slop for anything going out the door, since it assumes the first rewrite failed and scans again. Reach for voice-injection-rewriter when the output has to sound like a particular person, and for humanize-academic-writing when the register is scholarly and the ethics of the edit need stating in the file. Keep ai-writing-detection for the opposite question, the one about somebody else's text, where the cost of being confidently wrong lands on a person.

The trap this avoids is the one at the top: an agent editing on vibes, checking its own work by feel, and a file called humanize that turns out to be a shipping label rather than an editor. The name is the topic. The body is the skill.