Skip to content

forint573/interpreting-hungarian

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Interpreting Hungarian

English ↔ Hungarian translation skill for AI agents

"Ha már Mesterséges Intelligenciát használsz magyar szöveg készítésére, legalább tedd úgy, mint akinek van esze. (Mondod te a modellnek, melyet alkalmazol)"

Ez a SKILL Fable 5-el készült, összesen 15M token van mögötte, és még mindig jobb, ha te is belematatsz az AI generált szövegbe, de ezzel valószínűleg jobb eredményt érsz el mint nélküle. Még akár Fable 5 és GPT 5.6 esettében is. Sonnet 5, Opus 4.8, GLM 5.2, Kimi 2.7 és DeepSeek V4 simán megbírja és ezekkel a modellekkel is lehet közel végleges MI szöveget készíteni a SKILL alkalmazásával.

Translate English to Hungarian and Hungarian to English the way a native writes, not the way a dictionary maps words. A drop-in Agent Skill for Claude, Claude Code, Codex, and any agent that reads the Agent Skills format.

License Claude Skill Language pair Models

Hungarian is hard in both directions. A model's raw output usually parses — Hungarian translation studies calls the residue kvázi-helyesség: every sentence grammatical, the whole still unmistakably translated. Going the other way it produces Hunglish: English words in Hungarian word order, the tense system flattened, the articles missing. This skill encodes what closes both gaps — including the devices corpus research shows translations systematically lack, because the source language never prompts them.

What it catches

Traps a strong model walks into unaided — the middle column is the instinctive translation, and it is wrong:

English → Hungarian

English Instinctive but wrong Native
I've lived here for two years. Két éve itt laktam. (= two years ago) Két éve itt lakom.
I'm hot. Meleg vagyok. (= I'm gay) Melegem van.
You must not smoke here. Itt nem kell dohányozni. (= no need to) Itt nem szabad dohányozni.
Do you play tennis? Játszol teniszt? Teniszezel?
I like your dress. Szeretem a ruhádat. Tetszik a ruhád.
I can't help laughing. Nem tudok segíteni nevetve. Muszáj nevetnem.
Mistakes were made. Hibák lettek elkövetve. Hibákat követtek el.
"Come in," said Peter. "Gyere be," mondta Péter. – Gyere be – mondta Péter.
a $5 billion deal 5 billió dolláros üzlet (1000× off) 5 milliárd dolláros üzlet

Hungarian → English

Hungarian Instinctive but wrong Native
Két éve láttam. I've seen him for two years. I saw him two years ago.
Már két éve itt lakom. I live here since two years. I've lived here for two years.
Anna alszik, Péter pedig olvas. Anna is sleeping, Peter however is reading. Anna is sleeping, while Peter is reading.
PÉTER ette meg az almát. Peter ate the apple. (emphasis lost) It was Peter who ate the apple.
Senki nem segít. Nobody doesn't help. Nobody helps.
Öten voltunk. We were five. There were five of us.
Azt hiszem, hogy jó. I think, that it is good. I think it's good.
Nagy Éva felhívott. Nagy Éva called me. Éva Nagy called me.
ötmilliárd forint five milliard forints five billion forints

Beyond the traps, the skill covers the full system: topic–focus word order, definite vs. indefinite conjugation, coverb aspect, case suffixes, register (te/ön) and its English compensation, relative pronouns, false friends both ways, literary norms (dialogue dashes, no tense backshift, names, titles, form-faithful verse), domain conventions (software UI, Netflix-style subtitles, marketing, plain-language official prose), and punctuation/date/number conventions in both directions. It works as a diagnostic, not a checklist — it names over-correction as a failure mode, so already-natural sentences are left alone. And it restores what translated Hungarian statistically lacks (szokott, -lak/-lek, hadd, particles, diminutives): the source text never prompts them, natives use them anyway.

Install

Core skill in one line (self-contained for everyday English→Hungarian):

curl -fsSL https://raw.githubusercontent.com/forint573/interpreting-hungarian/main/interpreting-hungarian/SKILL.md --create-dirs -o ~/.claude/skills/interpreting-hungarian/SKILL.md

Full folder — recommended, and required for Hungarian→English, literary, and domain work:

git clone https://github.com/forint573/interpreting-hungarian.git
cp -r interpreting-hungarian/interpreting-hungarian ~/.claude/skills/
  • Per project: use .claude/skills/ inside the repo instead of ~/.claude/skills/.
  • Codex / Cursor / Gemini CLI: copy the same folder into that tool's skills directory (e.g. ~/.codex/skills/); for agents without skills support, paste SKILL.md into the system prompt or AGENTS.md.
  • Claude desktop & web apps: zip the interpreting-hungarian folder and upload it under Settings → Capabilities → Skills.
  • Claude API / Agent SDK: load the folder as an Agent Skill, or drop SKILL.md into the system prompt.

Use

Restart your session and just ask — the skill triggers on anything that looks like English↔Hungarian translation:

  • Translate this email to Hungarian: …
  • Fordítsd le angolra: …
  • Localize this landing page for a Hungarian audience.

It returns only the translation, preserves structure and formatting, leaves code, URLs, brands, and placeholders untouched, and keeps one register per text (domain defaults; formal ön when ambiguous — say informal (te) or formal (ön) to override). Going into English, the dead te/ön distinction is compensated with address forms, directness, and contractions.

Model notes

  • Claude Sonnet 5 / Claude Opus 4.8 — rules state their scope explicitly, reference files are routed with literal "read before drafting" triggers, and deliberation is directed only where it pays (register maps, names, terminology for long or literary passages). API: default effort: high suits everyday texts; xhigh pays off for poetry and full-document localization. Sonnet 5 takes no temperature — the skill's explicit conventions supply that consistency.
  • GPT-5.5 — written to match its literal instruction-following: absolutes are true invariants (never the kerül-passive, only the translation in the reply), everything else is a decision rule judged by outcome. A calibration section in SKILL.md addresses each model family by name. Reasoning medium is enough for everyday texts, high for literary.
  • Other models run it unchanged — positive patterns first, calque examples as labeled contrast, worked examples throughout.

Inside the folder

File What it holds
SKILL.md The diagnostic core: EN→HU rules, the HU→EN essentials, routing, output contract, self-checks
reference/hungarian-to-english.md Full HU→EN system: reassembly method, tense fan-out, articles, negation, names, register compensation, Hunglish sweep
reference/literary-translation.md The műfordítás norms: dialogue dashes, tense discipline, names/titles/realia, verse, transfer operations
reference/domain-conventions.md Software UI, subtitles (Netflix limits), marketing, plain-language official prose
reference/grammar-tables.md Full paradigms: conjugations, cases, postpositions, possession, vowel harmony
reference/style-lexicon.md Greetings, punctuation, dates/numbers, extended false friends, MT tell list

The core file works alone; the agent pulls reference files in only when the text type needs them, so the always-loaded footprint stays small.

FAQ

How do I translate English to Hungarian with Claude Code? Install the skill, then ask for the translation in plain language — it loads automatically, no flags or API keys.

Does it translate Hungarian to English too? Yes — install the full folder; reference/hungarian-to-english.md carries that direction.

Which models and tools? Tuned for Claude Sonnet 5, Claude Opus 4.8, and GPT-5.5; compatible with earlier models. Works in Claude Code, the Claude apps, Codex, Cursor, the Gemini CLI, and via the API.

Does it affect other languages? No — it only activates for English↔Hungarian requests. Free under Apache 2.0.

Grounding

The rules are drawn from reference grammars (Rounds; É. Kiss on information structure), the Hungarian orthographic standard (AkH. 12.), Klaudy Kinga's transfer-operation framework and translator-training methodology, corpus research on translation universals (the unique-items findings for Hungarian), the műfordítás canon from Arany to Nádasdy, and the professional style guides of the Hungarian localization industry (Microsoft, FSF.hu, Netflix, the EU institutions).

License

Apache License 2.0 · Copyright 2026 Virág Làzár Csaba ✸☽

Keywords

English to Hungarian translation, Hungarian to English translation, translate English to Hungarian, translate Hungarian to English, AI Hungarian translation, idiomatic Hungarian, native Hungarian translation, interpreting Hungarian, Hungarian interpreter AI, magyar fordítás, angol-magyar fordító, angol magyar fordító, angol magyar fordito, magyar-angol fordító, magyar angol fordító, magyar fordító, magyar fordito, műfordítás, Hungarian literary translation, Hungarian subtitles, Hungarian localization, Hungarian UI localization, Hunglish, leiterjakab, Claude skill, Claude Code skill, Claude Agent Skill, Codex skill, Claude Sonnet 5, Claude Opus 4.8, GPT-5.5, machine translation Hungarian, English Hungarian translator, Hungarian English translator.

About

English to Hungarian translation skill for Claude. Idiomatic, native Hungarian instead of literal calque, built on the most common EN to HU translation mistakes.

Topics

Resources

License

Code of conduct

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors