What is the stop-slop skill for Claude Code?

Claude Code writes a README, a changelog or a launch post, and a reader can tell quickly that a model wrote it. The stop-slop skill by hardikpandya is a short rulebook that makes the agent catch those tells before it hands the text back, and it has 16,862 installs in the Skillselion catalog (skills.sh registry, 2026-10-13).

Key takeaways

• stop-slop records 16,862 installs, 16,840 GitHub stars, quality tier A and a MEDIUM risk rating on its stop-slop skill listing (Skillselion catalog API, 2026-10-13).
• The repository itself shows 17,959 stars, 1,315 forks and an MIT license (GitHub API, 2026-10-13). The catalog star count is a lagging mirror of GitHub.
• The published SKILL.md holds 8 core rules, 12 quick checks and a 5-dimension score. A draft that scores below 35 out of 50 gets revised.
• It ranks third by installs among the 20 skills the catalog returns for "slop". The leader, unslop from cursor/plugins (22,852 installs), cleans generated code diffs. Among prose skills only no-ai-slop (21,651 installs) is ahead.
• It is an English prose skill. For Chinese text, humanizer-zh has 52,394 installs, more than three times as many.

What is the stop-slop skill for Claude Code?

stop-slop is an editing skill that tells Claude Code which phrases, sentence shapes and rhythms to remove from prose. The whole skill is one SKILL.md plus 3 reference files (phrases.md, structures.md and examples.md), so it adds no tools, no scripts and no MCP server. The catalog places it in Documentation, one of 62,347 skills indexed on 2026-10-13 (Skillselion catalog API).

The frontmatter says to use it when drafting, editing or reviewing text. Claude Code can load it without a slash command whenever a task involves writing.

What does stop-slop tell Claude Code to cut?

stop-slop gives the agent 8 core rules. In the order of the published SKILL.md:

1. Cut filler phrases, including openers that announce a point before making it.
2. Break formulaic structures such as binary contrasts and dramatic fragments.
3. Use active voice, with a human subject in every sentence.
4. Be specific. Name the thing in place of a vague declarative.
5. Put the reader in the room. "You" beats "people".
6. Vary rhythm. Mix sentence lengths and remove em dashes.
7. Trust readers. State facts without softening.
8. Cut quotables. A line that sounds like a pull quote gets rewritten.

The reference files make those rules concrete. phrases.md lists 15 adverbs to delete by name, among them "really", "actually" and "fundamentally", and a table of 11 business jargon swaps such as "deep dive" to "analysis". structures.md holds a table of 11 binary contrast rows, the "not X, it's Y" family, with one instruction: state Y.

How does the stop-slop scoring work?

stop-slop scores a draft from 1 to 10 on 5 dimensions, and anything below 35 out of 50 goes back for revision. The dimensions are directness, rhythm, trust, authenticity and density. Before scoring, the skill runs 12 quick checks, for example "Three consecutive sentences match length? Break one."

The model scores its own draft, so use the number to force a second pass. Ask Claude Code to print the 5 numbers and the checks that fired. That output shows you what it changed and why.

How do you install stop-slop in Claude Code?

Run npx skills add hardikpandya/stop-slop. The repo holds a single skill at its root, so no --skill flag is needed. The README also describes use outside Claude Code: upload SKILL.md and the reference files to a Claude Project, or include SKILL.md in an API system prompt.

After the install, ask for an edit in plain words: "Run stop-slop on docs/launch-post.md and show the score." The repository was last pushed on 2026-03-17 (GitHub API), so the rule set has been stable for about 7 months.

What are the best alternatives to stop-slop?

The closest prose alternatives to stop-slop (16,862 installs) are no-ai-slop, write and humanizer (skills.sh registry, 2026-10-13). Only no-ai-slop has more installs:

• no-ai-slop, a skill by petergyang: 21,651 installs, 7,277 stars.
• write from tw93/waza: 14,873 installs, 6,962 stars. It covers Chinese and English prose.
• humanizer, a skill by blader: 9,019 installs and 43,525 stars, built on the Wikipedia list of AI writing signs.

stop-slop has more GitHub stars than no-ai-slop: 16,840 against 7,277 in the catalog.

unslop, a skill from cursor/plugins has 22,852 installs but targets slop in generated code diffs. Do not use it for prose.

Another unslop, from poteto/noodle, has 780 installs and is covered in the poteto skills guide.

The best skills for docs and planning page ranks the wider category, and the guide to Claude skills to write like a human compares the editing skills side by side.

A complete starter stack

1. stop-slop for the phrase and structure pass on any English draft.
2. humanizer for a second opinion that uses a different pattern list.
3. A skill from the docs and planning category for the document type itself. documentation-writer from github/awesome-copilot has 27,312 installs (skills.sh registry, 2026-10-13).
4. Your own style guide in CLAUDE.md, so product terms survive the edit.

Common pitfalls

• Running it on code comments, configs or quotes. The rules target prose. Exclude quotes. An edit changes what the customer said.
• Treating "kill all adverbs" as law. The skill bans every -ly word. Technical writing sometimes needs "atomically" or "eventually", so review the diff and restore those.
• Stacking it with humanizer in one pass. Two rule sets with different ban lists make the agent rewrite the same sentence twice. Run them one after the other and compare.

Install it, run it on one real document, and keep it if the 50-point score and the diff agree with your own read.

Common questions