stop-ai-slop-jp: a Claude Skill that strips AI tells from Japanese prose

The problem stop-ai-slop-jp targets is register, not grammar

Machine-drafted Japanese is often clean at the sentence level. The README's complaint is elsewhere: headings that inflate a small observation into a revelation, a personal anecdote that suddenly climbs to 真理 or 美学, the same paragraph length and the same landing repeated down the page, leftover `**` markers, full-width dashes, decorative emoji. Each of those is locally harmless and collectively readable as machine output.

The project is a Claude Skill, meaning it is a set of instructions loaded into a Claude session rather than a program you run over a file. That choice matters for who it is for. It suits a writer who already has a draft and wants a second pass with named reasons, and it does not suit anyone who wants a build step that fails a pull request. The README is explicit that the skill cannot repair 書き手の不在, the absence of a writer, and points to writing in your own words as the shorter route out. That is an honest boundary and it also caps what the tool can do for you.

What the repository actually contains

The layout is small and flat. `SKILL.md` holds the core rules, a quick check and a scoring table. `references/phrases.md` lists vocabulary to remove: 偏愛語, translation-flavored constructions, hedges, foreign-script metaphors, emoji. `references/structures.md` covers habits that sit below vocabulary: false agency, propositional H2 headings, hook-then-reversal openings, absence of variation, and the both-sides summary. `references/examples.md` pairs an AI version with a corrected version.

There is no code, no test suite and no build. The only executable step in the whole project is cloning it. That has a consequence worth stating plainly: every judgement the skill makes is a judgement a language model makes while reading those markdown files. There is no deterministic rule engine behind the checks, so the same paragraph reviewed twice may not receive identical findings. The reference files are what constrain the model, and their quality is the ceiling on the skill's usefulness.

How it installs and how a first review runs

For personal use in Claude Code, the README gives a single clone into the skills directory. The target path is fixed, so the directory name must stay `stop-ai-slop-jp`.

git clone https://github.com/iKora128/stop-ai-slop-jp ~/.claude/skills/stop-ai-slop-jp

For a per-project install, the same clone goes under the project's own `.claude/skills` directory instead.

git clone https://github.com/iKora128/stop-ai-slop-jp <project>/.claude/skills/stop-ai-slop-jp

Two other delivery routes are documented. Claude Projects takes `SKILL.md` and the `references/` directory as uploaded project knowledge. Through the API, `SKILL.md` goes into the system prompt and the reference files are loaded as needed. The README does not describe how to wire the reference files in the API case beyond that phrase.

Once installed, the skill is invoked by asking for it by name. The README notes that naming it explicitly is more stable than hoping it triggers. A first review pass, with no rewriting, uses a prompt like this:

この文章を stop-ai-slop-jp の基準でレビューして。
重大な順に、立場、主体、構造、語彙、記号のどこが臭いか指摘して。
各指摘は「問題箇所 / なぜ臭いか / 直すならどうするか」で出して。
まだ書き換えなくていい。

The expected output is a list of findings ordered by severity, each with the offending passage, the reason it reads as machine-written, and a suggested repair. Nothing is rewritten in this mode, which makes it the right first step: you see the diagnosis before you accept the treatment.

The repair order is the real design decision

The README fixes a priority: 立場, then 主体, then 構造, then 語彙, then 記号. Stance first, agency second, structure third, vocabulary fourth, punctuation last. It states directly that fixing only dashes and emoji leaves the AI smell intact while 書き手の不在 remains.

That ordering is the part of the project worth arguing with, and I think it is right. Vocabulary substitution is the easiest edit to automate and the least likely to change how a paragraph reads. A draft whose every paragraph is the same length and ends on the same kind of summary will still read as generated after you swap out every 偏愛語. The scoring table encodes the same priorities across five axes: 立場 (is there a falsifiable specific claim), リズム (do length, tone and conclusions vary), 主体性 (is it clear who did what), 具体性 (does it descend to a specific context instead of stopping at abstractions), 削減 (is there anything left to cut). Each is scored 1 to 10, and the README's threshold is a rewrite below 35 out of 50.

Two of those axes are doing more work than the other three. 立場 and 主体性 are the ones a model cannot fake on your behalf, because they require you to have decided something. A skill can flag a sentence that has no agent. It cannot supply the agent.

Where it fails, and who should not use it

The honest limitation is in the README, not in a footnote: the skill cannot fix an absent writer. If the draft contains no position, no specific claim and no one doing anything, the review pass will produce a list of symptoms and the rewrite will produce tidier prose with the same emptiness. You will have moved the smell rather than removed it.

A second limitation follows from the delivery model. This is a Claude Skill. The README documents installation for Claude Code, Claude Projects and the API, and mentions no other client. If your team writes Japanese in a different assistant, or wants a check that runs on every commit, this is the wrong shape of tool: there is no CLI, no exit code, no configuration file to point at a directory of drafts.

A third is scale. The skill reviews a passage you paste into a conversation. Nothing in the repository describes batch processing a folder of articles, and the API route is described only as placing `SKILL.md` in the system prompt. If your problem is a hundred translated pages, the per-conversation workflow is the bottleneck.

Finally, the skill is tuned for Japanese. The patterns in `references/phrases.md` and `references/structures.md` are Japanese patterns, and the README presents the project as a Japanese edition of an English original. Using it on English text means running rules written for a different language.

The alternative is the English original, and the difference is not just translation

The README credits hardikpandya/stop-slop as the source of the idea, and the repository ships `README.en.md` alongside the Japanese one. The two projects share a shape: a skill file, reference lists of banned vocabulary and structures, and paired examples. The difference is the pattern inventory. The Japanese edition adds categories that exist because Japanese prose has them: 全角ダッシュ, unnecessary 「」 quotation marks, 翻訳調 constructions that read as translated rather than written, and the specific habit of inflating a small experience toward 真理 or 境地. The English original's lists would not name those.

If you write English, the original is the closer fit. If you write Japanese, the port is the one carrying the relevant patterns, and the credit line and the note article it links to are where the Japanese examples were assembled. The choice is not about which is better maintained; it is about which vocabulary list matches the language you are editing.

Maintenance, licence and what an upgrade costs you

The repository is not archived, and the last push was on 2026-06-11. There are no releases, which is consistent with a project that has nothing to build: distribution is a clone of the default branch, so there is no version number to pin and no changelog entry to diff against a tag. `CHANGELOG.md` exists at the top level, so changes are at least recorded there.

Upgrading is therefore a `git pull` in whichever directory you cloned into, and reviewing `CHANGELOG.md` and the reference files for new patterns. The cost is low in mechanical terms and non-zero in behavioural terms: because the reference files are the rules, any edit to them changes what the model flags on your next review. A new entry in `phrases.md` can start rejecting a construction you use deliberately.

The licence is MIT, stated in the README and shipped as `LICENSE`. MIT permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. The README also credits two external sources, the original repository and a note article by だいち | GENSHI AI, for the Japanese pattern work and examples. If you fork or redistribute, keeping those credit lines intact is the straightforward reading of the README's own クレジット section. This is a description of what the files say, not legal advice.

Editorial conclusion

Adopt it if you draft Japanese in Claude and currently publish without a review pass; the install is one git clone and the review prompt asks for findings before rewrites. Skip it if your source text is English, or if you want a linter that runs in CI, since nothing here executes outside a Claude session. Before relying on it, read references/examples.md and check whether the AI/human pairs match the register you actually write in, and confirm that your Claude client loads skills from ~/.claude/skills rather than only from project knowledge.

Frequently asked questions

Official sources

1. iKora128/stop-ai-slop-jp on GitHub
2. Issues
3. License: MIT
4. README