FontLab writing skills¶

Agent skills that write in the FontLab house voice. Install them into Claude Code, Cursor, Codex, or any tool that reads SKILL.md files:

npx skills add Fontlab/vexy-fontlab-writing-skills

Add one at a time with -s, or take the set:

npx skills add Fontlab/vexy-fontlab-writing-skills -s fontlab-neutral
npx skills add Fontlab/vexy-fontlab-writing-skills --all

Current localization guidance follows the core memories. Language pages state the current decisions; dated review ledgers and Git history retain earlier wording. Term tables are regenerated from the core memories, including their notes and approval status.

The skills¶

Skill	Use it for
`fontlab-neutral`	Release notes, what's new, announcements, introductions, About pages. The primary voice skill.
`fontlab-marketing`	Landing pages, product pages, store copy, campaigns, taglines, upgrade mail, comparison pages, offers and objections, search-ready product passages, AI-assisted marketing drafts.
`fontlab-technical`	Manuals, procedures, reference pages, help articles, tooltips, API docs.
`fontlab-write`	New text that connects emotional appeal to precise technical explanation.
`fontlab-rewrite`	Existing mixed text: recognize each passage's job, edit in the matching style, and smooth the transitions.
`fontlab-tldr`	Literary condensation to about 20%, preserving source voice, structure, named identities, and distinctive phrases through three private rounds.
`fontlab-simplify`	Plain-language, controlled-English (ASD-STE100 writing rules) or easy-to-read rewrites that keep every fact, condition, hedge and label.
`fontlab-terminology`	Product names, term checking, the nouns that mean different things in different apps.
`fontlab-localization`	Writing English that survives translation, Qt catalog mechanics, translation memories, machine drafts, and the error typology a review measures against.
`fontlab-localization-de`	German: register, headline-style compression, compounds and loanwords, plural and number facts, mnemonics, key names, false friends, and the German term table.
`fontlab-localization-es`	Latin American Spanish (`es_MX`): register, dialect discipline, terminology, plural and number facts, mnemonics, key names, and the Spanish term table.
`fontlab-localization-fr`	French: register, the Haralambous-based terminology, typographic spacing, plural and number facts, mnemonics, key names, and the French term table.
`fontlab-localization-pl`	Polish: register, case and gender with placeholders, the matryca and firet decisions, four plural categories against Qt's three forms, diacritics, and the Polish term table.
`fontlab-localization-zh`	Simplified Chinese (`zh_CN`): register without pronouns, verb-object labels, the single plural form with classifiers, mainland punctuation and vocabulary, appended mnemonics, false friends, and the Simplified Chinese term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-zh-hant`	Traditional Chinese (`zh_TW`): register, Taiwan's punctuation and character forms, the single plural form with classifiers, appended mnemonics, false friends, and the Traditional Chinese term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-ru`	Russian (`ru_RU`): the вы register, sentence case, three plural forms, mnemonics on Cyrillic letters, alphabet mixing, false friends, and the Russian term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-pt`	Brazilian Portuguese (`pt_BR`): the você register, the 1990 orthography, two plural forms under the `pt_BR` rule, mnemonics on unaccented letters, false friends, and the Portuguese term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-ar`	Arabic (`ar`): Modern Standard Arabic, verbal-noun commands, six plural forms, mnemonics on Arabic letters, bidirectional rules, false friends, and the Arabic term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-hi`	Hindi (`hi_IN`): the आप register, two plural forms, appended mnemonics, Devanagari encoding and spelling, false friends, and the Hindi term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-ja`	Japanese (`ja_JP`): the です・ます register, counters and the single plural form, Japanese–Latin spacing, appended mnemonics, katakana loans, false friends, and the Japanese term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-it`	Italian (`it_IT`): the tu register, two plural forms and their agreement, mnemonics on unaccented letters, names and loans, false friends, and the Italian term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-id`	Indonesian (`id_ID`): the Anda register, the single plural form without reduplication, mnemonics, names and loans, false friends, and the Indonesian term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-ko`	Korean (`ko_KR`): the 하세요 register, counters and the single plural form, appended mnemonics, Hangul-only spelling, false friends, and the Korean term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-tr`	Turkish (`tr_TR`): the siz register, the single plural form with singular nouns after numbers, the dotted and dotless i, mnemonics, false friends, and the Turkish term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-vi`	Vietnamese (`vi_VN`): the bạn register, the single plural form, mnemonics on unmarked letters, tone-mark placement, false friends, and the Vietnamese term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-th`	Thai (`th_TH`): register without pronouns or particles, classifiers and the single plural form, appended mnemonics, line breaking and character order, false friends, and the Thai term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-uk`	Ukrainian (`uk_UA`): the ви register, the 2019 orthography, three plural forms, mnemonics on Cyrillic letters, the apostrophe letter, false friends, and the Ukrainian term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-localization-cs`	Czech (`cs_CZ`): vykání, sentence case, three plural forms, mnemonics on a letter of the translation, false friends, and the Czech term table. Machine-drafted guidance and proposed term table; no native review yet.
`fontlab-partners`	Content for partners.fontlab.com, where the traps are structural rather than stylistic.

Each skill is self-contained. It installs on its own, carries every rule it enforces, and reads no file outside its own directory.

Balanced is a deliberate mixture, not a new name for neutral. Use neutral for the established factual house voice. Use balanced writing when a new piece needs both reader appeal and technical detail; use balanced editing when a draft already mixes those jobs. There is no fixed percentage, and limitations, prices, and procedures stay factual. The balanced skills reuse the measured registers; no separate balanced corpus or score is claimed.

Use /fontlab-tldr to condense a source while keeping its own narrative voice. It combines neutral-writing discipline with three rounds at the same 20% target, then returns only the final text. When no distinctive English style is identifiable, it uses ASD-STE100 as the fallback.

Prompts for tools without skills¶

Not every chat installs skills. The prompts carry the same rules for copying and pasting, one page per job: neutral text, marketing and technical writing or editing, balanced writing or editing, TLDR and a review prompt for a second reader. Each page has a short variant and a long one. The long variant is generated from the skill by tools/scripts/sync_prompts.py, so it cannot drift; the short variant is maintained by hand.

How the skills make their rules hold¶

A long rule set asks the agent to remember everything while it writes, and models drop instructions as their number grows. So each skill starts with the order of work and ends with a checklist of eight to twelve yes-or-no questions about the text and its evidence, tagged with the house rule each enforces. House rule H17 says how to use it: after every portion of the work, answer each item separately, quote the words of any failure, repair only what fails, recheck the repairs, mark an item that fails twice and stop after three rounds, marking what still fails. Where the agent can open a subagent or a fresh chat, it hands that reviewer the draft, the evidence and the checklist, not its own reasoning. fontlab-tldr runs the checklist inside its three rounds and adds none.

The writing guide's For agents page gives the research behind this.

How the skills use the house voice¶

Neutral states facts and changes warmly. Marketing connects an offer with the reader's work. Technical writing explains a mechanism or gives usable instructions. Choose the register per passage; do not force a whole document into one register because it contains a procedure or an offer.

The shared craft rules teach concrete attention, connected paragraphs and varied pace. A developed sentence can explain a relationship; a short one can settle it. Recurring details should acquire meaning as the explanation proceeds. The neutral skill demonstrates this method through paired short and patient examples. The remaining skill cores are being revised individually.

The skills preserve useful conditionals, qualifications, rhythm, and exact source material. They do not add pronouns, named tools, metaphors, or punctuation to meet a quota. A corpus measurement can identify a passage worth inspecting; it cannot establish correctness, usability, or authorship.

Worked cases state their evidence before the revision. Fictional examples are labelled and do not establish real product behavior or current offer terms. Every skill carries the shared factual, editorial, and output safeguards.

The measurement tool remains an optional diagnostic aid. Its stored register bands are legacy comparison data, not writing requirements. Below 40 words it omits range comparisons; this display threshold does not make longer samples reliable. Word matches are review candidates, including accurate technical uses. Its lightweight Markdown cleanup is approximate: select coherent prose and inspect quotations, labels and examples separately. It cannot verify facts or identify an author.

Maintaining them¶

The shared block lives once, in tools/house-rules.md, and is copied into each skill between markers. Never edit it inside a skill.

python3 tools/scripts/new_language_skills.py  # rewrite the fifteen machine-drafted language skills from the guide
python3 tools/scripts/export_terms.py         # regenerate the per-language term tables from ../vexy-fontlab-writing-styleguide
python3 tools/scripts/sync_shared.py          # write the block into every skill
python3 tools/scripts/sync_shared.py --check  # fail if any skill is out of date
python3 tools/scripts/sync_prompts.py         # regenerate the long variant of every prompt
python3 tools/scripts/check_paths.py          # fail on an unresolvable path or sibling name
bash tools/scripts/check_all.sh               # all of the above

The German, Spanish, French and Polish skills are hand-written and reviewed. The other fifteen language skills are generated. new_language_skills.py reads each language's localization guide in ../vexy-fontlab-writing-styleguide and writes the skill: a statement of the draft status, the guide's sections before "Current terminology", and a pointer to the term table. It turns relative links into plain text, keeps the house-rules block already in the file, and refuses de, es, fr and pl. To change a generated skill, edit the guide and run the script again. Pass language codes to limit a run, as with export_terms.py. Once a native reviewer takes a language over, move its code to the script's refused list and maintain the skill by hand.

The combined check also runs the script tests. Punctuation choices need editorial review; a dash alone does not fail the structural checks.

Website¶

./build.py (a self-contained uv script) renders this README and every skill with ProperDocs, MaterialX and the fltheme26 theme into docs/, published at https://fontlab.dev/vexy-fontlab-writing-skills/. The Site GitHub workflow rebuilds it on every push and commits the result. ./build.py --serve previews it locally.

Licence¶

MIT. See LICENSE.