Writing Skills: Ten You Can Actually Install

The shortlist first: ten rows, install commands below

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

# 1. stop-slop — a single file, the lightest item here
mkdir -p ~/.claude/skills/stop-slop
curl -o ~/.claude/skills/stop-slop/SKILL.md \
  https://raw.githubusercontent.com/boraoztunc/skills/main/stop-slop/SKILL.md
# Same repo, other skills: swap the directory name. Don't clone the whole thing.

# 2. blader/humanizer — invoke with /humanizer once installed
npx skills add blader/humanizer --global

# 3. Aboudjem/humanizer-skill — the only one with a recomputable score
npx skills add Aboudjem/humanizer-skill

# 5, 6. Both official ones ship in the example-skills plugin
/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills

# 8. The marketing bundle (note: this pulls in 50 skills at once)
npx skills add coreyhaines31/marketingskills

# 9. Full SEO suite
/plugin marketplace add AgriciDaniel/claude-seo
/plugin install claude-seo@agricidaniel-claude-seo
/seo setup

# 10. Two pipelines — non-fiction and fiction
npx skills add nadiem99/claude-writing-skills
/plugin marketplace add haowjy/creative-writing-skills

# Do these two things immediately after installing
/skills          # review the list, disable what this project won't use
/skill-doctor    # token cost and invocation counts per skill

Item by item, plainly

• 1. stop-slop: the description is one line — "Remove AI writing patterns from prose." Its value is that it lists what to delete in seven classes: throat-clearing openers ("Here's the thing:"), emphasis crutches ("Full stop," "Let that sink in"), business jargon (unpack / lean into / deep dive / circle back), adverbs (every -ly plus really / just / literally), meta-commentary ("as we'll see"), false agency ("the decision emerges"), and binary contrasts ("not because X, but because Y"). The file shows before/after rewrites. One file, one curl, the lightest thing here.
• 2. blader/humanizer: 49.4k stars per its GitHub repo page, 2026-09, the best-known of the bunch. Twenty-five patterns in five groups (staging instead of stating / rhythm by rule / inflation and borrowed authority / formatting by rule / leftover draft elements). Invoke it with /humanizer; the output comes in three parts — rewrite, self-critique, final. Its framing is worth keeping: a language model writes whatever is most likely to come next, not a sentence it meant. It states plainly that it never invents facts — the one promise a rewriting skill owes you.
• 3. Aboudjem/humanizer-skill: 245 stars per its GitHub repo page, 2026-09. The longest pattern list (55 across five categories) and five voices (casual / professional / technical / warm / blunt), but what makes it unique here is a 0-100 AI-tell score plus a dependency-free Node metrics CLI: it scores text and diffs versions. It also fact-checks the rewrite against the original for numbers, URLs, dates and acronyms. MIT, no telemetry, nothing leaves the machine.
• 4. humanize-writing: the listing on the Agensi marketplace; Skill 17 digs into this one. The point to take here is that "humanize / humanizer / humanize-writing" is not a skill's name, it's a genre's name. Findable implementations include blader, Aboudjem, gencharitaci/humanize-skills, and entries from lguz, jpeggdev and shir-danishyar on skills.sh. Check which repository you are installing.
• 5. doc-coauthoring: official, and the only item here that bundles no scripts and no reference files at all. It gives three stages: info-dump your background to the model (team discussions, architecture details, stakeholder concerns), then work section by section through "ask → brainstorm 5-20 options → curate → check gaps → draft → refine," then reader-test by handing it to a subagent or a fresh Claude instance and watching where comprehension breaks. It treats a different disease than prose quality: the context still stuck in your head.
• 6. internal-comms: four format files — examples/3p-updates.md (progress / plans / problems), company-newsletter.md, faq-answers.md, general-comms.md — read as needed. Its description is written in first person: write internal comms "using the formats that my company likes to use." That line is the usage note: fork it and swap the examples for your own, don't install it as-is.
• 7. better-writing: the name misleads. This is UX copy — button labels, form errors, placeholders, settings, empty states. All twelve rules are checkable: start buttons with a verb ("Save draft," not "OK"), label toggles for their ON state rather than the negative, put error text beside the failing field and say what to do, use one vocabulary across a multi-step flow, make link text meaningful out of context. This is the closest thing here to a writing skill with objective criteria — interface copy can be judged against the interaction itself.
• 8. copywriting · copy-editing: from coreyhaines31/marketingskills (50.7k stars and 7.7k forks per its GitHub repo page, 2026-09). The first writes landing and home page copy, the second reworks existing copy. The same repo carries cold-email, emails, social, marketing-psychology and about fifty directories in total. Which is also the problem: one npx skills add pulls in fifty descriptions, all competing for Claude Code's 1,536-character description budget. If you want two, take two.
• 9. seo-audit · programmatic-seo · claude-seo: the first two live in boraoztunc/skills (347 stars per its GitHub repo page, 2026-09), alongside schema-markup and content-strategy. For the full suite there's AgriciDaniel/claude-seo (17.1k stars per its GitHub repo page, 2026-09): 25 sub-skills and 18 subagents spanning technical SEO, schema, GEO/AEO, backlinks and reporting. The core needs zero API keys; PageSpeed Insights, Search Console and GA4 are optional enrichment, and DataForSEO-class extensions want your own accounts. It is the odd one out here — section 02 explains why.
• 10. Two pipelines: nadiem99/claude-writing-skills (2 stars per its GitHub repo page, 2026-09) breaks one Substack publication's editorial process into eight stages — interview → outline → draft → coach → edit → source-check → top-edit → repurpose — plus setup / research / status as support. Low stars don't mean not worth reading: it's the most literal example of turning your own process into skills, and Skill 33 returns to it. On the fiction side, haowjy/creative-writing-skills (468 stars per its GitHub repo page, 2026-09) uses story-memory, character-sim and reader-sim for continuity, character consistency and reader reaction — the hard part of a long serial is remembering chapter three, not choosing words.

Why these ten don't compare: four different things in one table

Four shapes

1. Subtractive (1-4) — text in, rewritten text out. The shared structure is a list of what to delete, running from 7 classes to 55 entries. This is the only branch of writing skills that can be checked mechanically: to see whether it fired, count the -ly endings and the em dashes. Item 3 simply turned that into a score.
2. Process (5, 10) — these don't tell you how to write a sentence, they tell you what to do in what order. The reader-test in item 5 and source-check in item 10 are the same move: forcing an external checkpoint into a subjective task.
3. Taste (6, 7, 8) — you're installing a person's or a company's preferences. Item 6 is Anthropic's internal comms format, item 8 is one marketer's copy doctrine, item 7 is a UI copy convention. This branch should be forked and replaced, never installed as-is. (Skill 15 hit the same trap with brand-guidelines.)
4. Technical checks (9) — filed under "writing," but it doesn't belong to this category: whether schema validates, what the page speed score is, whether a title exceeds its length — all objective. It behaves like the office category from Skill 13, and should be judged by those criteria.

Where the crack runs

Lay the four shapes side by side and it's clear: the more mechanically checkable, the less contested. Nobody argues about item 9, rarely about item 7; items 1-3 argue about deleting too much versus too little; items 6 and 8 argue about whose taste gets to win. The disagreement in this category isn't a quality problem, it's a property of the category — when the acceptance criterion is "I think so," install counts only measure whose marketing worked.

Three-minute inspection: six criteria, measured against these ten

1. Is what it deletes or changes written as a list? Item 1 lists seven classes with before/after pairs, item 2 lists 25, item 3 lists 55. Anything promising "more natural, flowing prose" without a list should be skipped — that's your own sentence handed back in different words.
2. Is there a recomputable signal? Item 3's 0-100 score and local CLI are the only case here of turning subjective acceptance into a number. Having a score doesn't make the score right, but it makes "did this change actually help?" comparable — the same requirement Skill 07 imposed on evals.
3. Does it promise not to touch facts? Item 2 states it never invents facts; item 3 explicitly checks numbers, URLs, dates and acronyms. The most expensive bug in rewriting is "23%" polished into "nearly a quarter" — smoother, and a different number. If SKILL.md doesn't make that promise, keep it away from anything with data in it.
4. Process or someone's taste? One glance at the description settles it: first person ("my company," "the formats I like") means template, not product — item 6 is the textbook case. Templates get forked and replaced; installing the original hangs someone else's voice under your name.
5. Star counts are least trustworthy here. Item 10's non-fiction pipeline has 2 stars and more substance than plenty of four-figure "writing assistants"; jooray/humanizer and blader/humanizer differ a thousandfold in stars and are the same thing. With no run result to vouch for anything, stars are pure reach. Go by the repo directory name and the body of SKILL.md.
6. Does it collide with what you already have? Items 1-4 compete for one trigger surface: the model picks one each time, and you don't control which. Keep exactly one per job — this matters more in writing than anywhere else, because the collision doesn't throw an error, it just makes your voice different every time.

The honest costs: what the bill looks like

No oracle means no reason to stop

Office skills have the luxury: if the file won't open it's wrong, and the model fixes itself. Writing has nothing like it. You will find yourself asking for one more version, round after round, with no version any closer to anything — because "done" is undefined. That's the real cost of this category, and it's why Aboudjem/humanizer-skill's 0-100 score is worth something even imperfect: it draws a stopping line.

Subtractive skills delete structure you needed

The "not X but Y," the -ly adverbs and the meta-commentary that stop-slop removes are often precision itself in a spec, an API doc or a contract clause. In "this parameter applies only in async mode," the word "only" is not filler; delete it and the meaning changes. Subtractive skills fit prose and marketing copy, not technical writing — and the trigger mechanism does not know the difference, so you have to.

The rest of the bill

• The description budget really does overflow. Item 8's marketingskills installs 50 skills at once; item 9's claude-seo carries 25 sub-skills. Claude Code merges all descriptions and truncates at 1,536 characters, after which the ones at the back can never fire — and it doesn't warn you, it just quietly stops working. If you want two skills, copy two SKILL.md files.
• Same-name forks are a supply-chain problem. A skill is executable instruction written by someone else, and this category has so many identically-named implementations that search cannot separate them. "Installed the one that looked like it" is the norm here, not the accident. Skill 30 covers vetting properly; for now: check whether the repo is a fork and when it last shipped a commit.
• De-slopping is an arms race. Detectors and rewriters chase each other; a phrase list that works this year may not next. The lists in items 1-3 are snapshots of an observation, not linguistic law. Skill 17 opens that pit on its own.
• Taste skills assimilate you quietly. Write landing pages with copywriting for three months and you'll start treating that author's doctrine as universal truth. Read this kind once as reference and rewrite the parts you agree with as your own rules, rather than leaving it on the trigger chain making judgments for you.