CommunityWriting & Editinggithub.com

prose

- Write and repair prose in German and English so it is free of errors and does not read as machine-written. Holds one opinionated style: no em dashes anywhere, the spaced en dash kept as the German Gedankenstrich but capped, emoji as literal section labels in headings only and never in body text, DIN 5008 typography in German, American spelling in English. Bundles a checker that masks the markdown and runs LanguageTool when a backend is chosen, so correctness is verified rather than claimed. Use this whenever text is being produced or edited for a human reader: notes, daily notes, documentation, READMEs, mails, report entries, the body of a commit message, comments in source files, or when the user pastes text and asks to clean it up, proofread it, fix the German, kill the em dashes, or make it sound less like AI. Use it even when the request says nothing about style, because output otherwise drifts straight back to em dashes and machine rhythm.

What is prose?

prose is a Claude Code agent skill that - Write and repair prose in German and English so it is free of errors and does not read as machine-written. Holds one opinionated style: no em dashes anywhere, the spaced en dash kept as the German Gedankenstrich but capped, emoji as literal section labels in headings only and never in body text, DIN 5008 typography in German, American spelling in English. Bundles a checker that masks the markdown and runs LanguageTool when a backend is chosen, so correctness is verified rather than claimed. Use this whenever text is being produced or edited for a human reader: notes, daily notes, documentation, READMEs, mails, report entries, the body of a commit message, comments in source files, or when the user pastes text and asks to clean it up, proofread it, fix the German, kill the em dashes, or make it sound less like AI. Use it even when the request says nothing about style, because output otherwise drifts straight back to em dashes and machine rhythm.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/nemesis398/prose/tree/HEAD/skills/prose

Installed? Explore more Writing & Editing skills: steipete/notion, langchain-ai/langchain, bytedance/podcast-generation · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

What does prose do?

Two jobs, same rules.

  • Compose. Write the text, then verify it, then hand it over.
  • Repair. Verify first, read what came back, fix, verify again.

The order matters in both directions. Composing without verifying means shipping em dashes you did not notice writing. Repairing without verifying first means guessing at what is wrong.

The loop

Nothing is finished until the checker has run on the actual saved text and you have read its output.

"${CLAUDE_SKILL_DIR}/bin/prose-check" path/to/file.md

If ${CLAUDE_SKILL_DIR} above was not filled in, the checker is bin/prose-check in the folder that holds this file (bin/prose-check.exe on Windows).

When the binary is missing, after a fresh install or a plugin update (an update lands in a new folder), run /prose:setup, or the install script beside this file (install.ps1 on Windows). The script reads the version from the plugin manifest two folders up; a copied folder has none, so set PROSE_VERSION to the release to install: a release tag of github.com/nemesis398/prose without its v. Until the checker has run, the text is unchecked: say so, and never report it as checked.

Useful flags: --no-lt skips LanguageTool and runs the rule pass only, --picky turns on LanguageTool's formal-text rules for something that has to be correct in front of other people, --json for parsing, --stdin for text that is not in a file yet, --strict to fail on warnings too, --selftest to see the settings and which backend answers.

Exit 0 means clean, 1 means an error was found. Exit 2 means the check did not run, never that the text is clean. Read the line it printed:

  • It names the settings file: the file is broken. Pass on what it says. /prose:setup repairs a wrong backend, server, username, rbw item or language rule. A file that is not valid JSON, or a wrong-typed key that setup does not write, such as ignoreWords, has to be fixed by hand.
  • It says LanguageTool did not run and names what is missing, such as the Premium account or key: run again with --no-lt, which still checks everything else, and say that LanguageTool did not run.
  • It names a backend and an HTTP status: the configured LanguageTool refused or did not answer. Say so. No other backend stands in for it.
  • It names an input: not a file: PATH, PATH is not UTF-8 (or <stdin> is not UTF-8) or cannot read PATH. The run stops before it checks anything, even when only one of several inputs is at fault, so no input was checked. Correct the path or the encoding, and run again.
  • It is a usage error: an unknown option, a missing or wrong --lang value, or no file at all (that prints the usage text). Correct the command line. No text was checked.

LanguageTool is off in the report means grammar was not checked. No backend is chosen, so nothing left the machine: the rules ran, and English spelling was checked by hunspell at most, when it is installed. Report that instead of a clean text, and point at /prose:setup to choose a backend.

--json prints one object per file, with that file's findings nested under findings. Flatten it with [x for f in json.load(sys.stdin) for x in f['findings']] and index x['line'], not .get: a filter over the top level matches nothing and reports a clean zero for a file with 16 errors.

Fix every error. Read every warning and decide, because some of them are judgment calls the checker deliberately leaves open. Then run it again. Report what the final run said rather than asserting the text is clean.

One error is not about the text at all. LT-CHUNK-FAILED means part of the document was never checked, usually a rate limit or a network problem. Re-run it, and never report a clean result from a run containing it: the unchecked stretch is precisely where an error would go unnoticed.

A huge count is a reason to read before fixing. When one rule fires hundreds of times in one or two files, check whether the findings cluster inside markup (links, HTML, inline CSS). That is a masking bug in the checker, not a discovery about the prose: five such bugs once invented 654 of 890 errors, and obeying them would have rewritten CSS to 17 px and broken two diagrams.

The checker covers the mechanical layer only. It cannot tell you that a paragraph restates itself or that every bullet has the same shape. A clean run is necessary, not sufficient.

Dashes

The em dash — is never correct here, in either language, in headings, in tables, in commit messages, in comments. It is the single clearest fingerprint of generated text.

German keeps the spaced en dash –. It is the correct Gedankenstrich, and careful German writers already use it. It is capped at roughly one per 200 words, and the checker reports the density for the document as a whole. The cap exists because the problem was never the character. The problem is reaching for the same move every third sentence, so the dash stops being a device and becomes the default rhythm.

English restructures. A comma, a colon, a parenthesis, or two sentences.

bad     Umami runs locally — no data protection agreement needed.
good    Umami runs locally, so no data protection agreement is needed.
good    Umami runs locally. No data protection agreement needed.

bad     Umami läuft lokal — kein AVV nötig.
good    Umami läuft lokal – kein AVV nötig.

Ranges keep the unspaced en dash in both languages: 2020–2024, 10–40 %, Mittel–hoch, S. 341–356.

Emoji

Emoji are labels, not decoration. That is the whole distinction, and it is what separates this convention from the way generated text uses them.

  • One per heading, as the first character, single space after.
  • Literal, matched to the content. 🍪 for the cookie banner, 🐞 for error tracking, 📊 for analytics, 🏫 for school. Never 🚀 for getting started, never ✨ for features, never 🎯 for goals.
  • Only on headings a reader scans for. Sub-headings that exist as internal scaffolding get none.
  • Never the same emoji twice in one document. A landmark that repeats is not a landmark.
  • None in body text, tables or bullets.
## 📊 Analytics          section anchor, gets one
### Vergleich            internal scaffolding, none
#### 🍪 Cookie-Banner    scan target, gets one
##### Empfehlung         internal, none

The exception is a document that already runs a legend, the way a reference note may mark its rows with the same few emoji throughout. Follow the convention already in the file instead of stripping it. The checker recognizes this: an emoji that repeats, or that also labels a heading, is reported as a legend and left alone.

Heading case

In English, Title Case for the note's name and its # heading, sentence case for every heading inside it. Proper nouns keep their capitals either way, which is why ## ☁️ Tier 2: restic to B2 Cloud Storage is already correct: restic is lowercase because that is the tool's name, and Backblaze sells a product actually called B2 Cloud Storage.

Backup Plan.md                          filename, Title Case
# Backup Plan                           H1, Title Case
## 🔢 3-2-1 in one paragraph            H2, sentence case
## 🧪 Open: none of this is restore-tested

German needs no rule here, because German capitalizes nouns regardless.

This one is deliberately not in the checker. Telling a wrongly capitalized heading from a run of proper nouns needs judgment the checker does not have, and a rule that fires on Proxmox Backup Server would be worse than no rule.

Voice

The mechanical rules above are enforced by the checker. This part is not, and it is where the actual work is. A banned-word list produces text that dodges the words and keeps the tic, because the tell is structural.

Vary sentence length on purpose. Uniform length is the strongest single signal of generated prose, because every sentence gets written to the same target. A four-word sentence after a thirty-word one is what a person sounds like.

Commit to a sentence before writing it. The dash habit is a symptom: the sentence finished, something else occurred to the writer, and a dash glued it on. Decide what the sentence is about first, then the punctuation follows.

Delete the sentence that restates the paragraph. Sections that end by summarizing themselves add length and no information. If the point was made, it was made.

Let bullets be ragged. If every bullet has the same length and the same grammatical shape, it was generated rather than observed. Real notes have a two-word bullet next to a three-line one, because reality is uneven.

Bold is a landmark, not emphasis. More than roughly one bold phrase per paragraph and nothing stands out any more.

Two or four, not always three. The rule of three is real, and reaching for it every single time is a tell.

Say the thing instead of announcing it. Cut let's look at the options and give the options. Cut it's worth noting and note it.

Prefer the verb to the noun, especially in German. Wir migrieren am Freitag beats Die Durchführung der Migration erfolgt am Freitag.

Name who does it. Passive hides the actor, and in a work note the actor is usually the useful part.

Be concrete. Numbers, names and dates instead of "significant" and "umfangreich". The register to match looks like this: 2–4 h, 5.000 Fehler/Monat, ab 35 €/Monat, 14 Monate Aufbewahrung.

The same goes for the shape of a note: terse, decision-first, tables for comparisons, bold on the decision itself, → for something superseded. Do not inflate a note into an essay.

Language

Pick the language from the text and the reader, not from the conversation. An existing text stays in its own language, with a language rule or without one.

Before writing, run "${CLAUDE_SKILL_DIR}/bin/prose-check" --rules and follow the language rule it prints: which language a text is written in, and when a German reader gets Sie. Without a rule, write new text in the language of the request, and address a German reader with lowercase du, held consistently through a document.

If --rules exits 2, or the binary is missing, the rule was not read. For a missing binary, run /prose:setup or the install script, as under The loop. For exit 2, the line it printed names the settings file; /prose:setup repairs a wrong languageRule, and the list above says which other faults it cannot. Do not guess a rule. Write as you would without one, and tell the user that their rule was not applied.

Mixed files are normal and correct. English headings over German prose in one note is a real pattern, not a mistake to clean up. The checker detects language per paragraph and sends each stretch to LanguageTool under its own language, so leave the mixture alone.

English is American spelling: organize, behavior, color, center.

References

Load these when the work is actually in that language. They carry the detail that does not belong in this file.

  • references/german.md for German. Komma rules that matter, das/dass, ss/ß, Groß- und Kleinschreibung, Komposita and the Deppenleerzeichen, numbers and units per DIN 5008, quotation marks, and the German phrasings that read as machine-written.
  • references/english.md for English. The dash policy in detail, the structural AI tells, and a section on German interference errors that survive proofreading because they are real English words. Read this one even when the English looks fine, because that is exactly when those slip through.
  • references/languagetool.md for the checker's settings file, its backends and key sources, self-hosting, and what to do when it will not connect.

Scope

This governs prose written for a human reader. It applies to files, to the body of a commit message, and to comments and identifiers in source.

In source files, apply the spelling and dash rules and nothing else. The emoji convention and the register guidance are for documents, not for code. For the shape of a commit message, conventional-commits decides; this skill only makes the prose inside it correct.

It does not govern chat replies.

Related Skills