Article Translator

A workflow for translating long-form prose with fidelity, consistency, and a preserved authorial voice. Translation here is treated as a series of deliberate choices over a whole text — not a sentence-by-sentence transcoding.

When to use

Input is more than ~3 paragraphs of prose bound for another language. Out of scope: UI copy, code, single sentences, tweet-sized fragments, song lyrics, metered poetry (verse translation needs a different toolkit), legal contracts (specialized terminology and disclaimers).

Core workflow at a glance

1. Analyze the source: language, locale, register, voice, purpose, and a brief structural outline of the argument.
2. Build a glossary of names, terms, numbers, dates, and recurring phrases — fixed renderings, applied throughout.
3. Decide strategy on idioms, units, punctuation, and format preservation.
4. Translate in paragraph-sized semantic units (Phase 4).
5. Footnote sparingly — only for preserved-but-opaque expressions or load-bearing ambiguities.
6. Self-review against the criteria below: drift, omissions, additions, naturalness, register, style.

Priority order on conflicts: accuracy → terminology consistency → naturalness → style preservation → pragmatic equivalence → cultural adaptation → skopos-fit.

Calibrate depth to the text. A 300-word blog snippet does not need a full glossary-and-review pass; a 3000-word longread for publication does. Skip nothing in spirit, but compress phases proportionally for short or low-stakes inputs.

Default output target: a separate file next to the original (see "File output" below). The chat reply summarizes; it does not paste the translation back.

Guiding criteria

The output should satisfy all of these criteria together. They are listed in priority order — when two conflict, the higher-listed criterion wins, and the choice should be made consciously.

1. Accuracy. Every proposition in the source must be present in the target. Nothing added, nothing dropped. Implications, hedges, modality, and irony all count as meaning that must survive.
2. Terminological consistency. The same concept gets the same word throughout, governed by the glossary built in Phase 2. Drift is the most common defect in long-form translation, and it is the easiest to verify.
3. Naturalness. The result must read as if originally written in the target language. If a native reader would never phrase it that way, rewrite. "Translator-ese" is a failure mode, not a neutral default — and the fix for it is bolder rewriting at the sentence and clause level, not closer adherence to the source.
4. Style and register preservation. Academic stays academic, casual stays casual. Sentence rhythm, signature devices (anaphora, parentheticals, fragments, em-dashes), and the author's idiolect must survive. Resist the urge to "clean up" deliberate quirks.
5. Pragmatic equivalence. Produce the same effect on the target reader as the original produced on the source reader. A joke must remain funny; a polemic must remain pointed.
6. Cultural appropriateness. Handle realia, allusions, units, and idioms with a deliberate, consistent strategy (see Phase 3).
7. Fit-to-purpose (skopos). A translation for publication, for personal understanding, and for an internal memo are graded on different curves.

When naturalness and literal accuracy conflict, naturalness usually wins — unless the user has explicitly asked for a literal or scholarly rendering. When terminology consistency and naturalness conflict, prefer consistency for nouns and domain terms; allow variation for pronouns and connective tissue.

Workflow

Translation moves through six phases: analyze → glossary → strategy → translate → footnotes → review. Do not skip phases. For very short inputs the phases can be compressed but never collapsed entirely.

Phase 1 — Analyze the source

Before writing a single word of translation, read the whole text and note:

• Source language, including any dialect or regional markers.
• Target language and locale (en-US vs en-GB, pt-BR vs pt-PT, zh-CN vs zh-TW, etc.). Confirm with the user if ambiguous.
• Genre and subgenre — op-ed, technical tutorial, personal essay, news report, marketing copy, etc.
• Register — formal, neutral, conversational, slangy, academic.
• Authorial voice and signature devices. Be specific: "frequent em-dashed asides", "very short paragraphs", "favors latinate vocabulary", "drops articles for punch", "alternates long analytical sentences with one-line zingers". These notes will steer Phase 4.
• Intended audience of the original.
• Purpose / skopos — what is this text trying to do (inform, persuade, entertain, sell, document)?
• Cultural anchoring — is the text steeped in a specific time and place? How transparent will its references be to the target audience?
• Structural outline. Briefly map the argument: main thesis in one sentence, the role of each section (setup, evidence, counter-argument, payoff), and any recurring motifs or callbacks. This map is for internal use during Phase 4 — it prevents losing the logical thread when translating section by section.

If anything about target locale, audience, or purpose is genuinely ambiguous, ask the user one focused question before proceeding. Do not ask just to seem thorough.

Phase 2 — Build the glossary

Before any translation, scan the full source and extract a glossary. The glossary is the single most effective defense against drift — the most common quality defect in long-form translation.

Default behavior:

• Texts >800 words with recurring proper nouns, technical terms, or leitmotifs → build a glossary.
• Texts <800 words with only a handful of unique terms → light glossary inline (just the few items that matter), or skip if the text has no repeating terms at all.
• Any text where the user asks for a quick rendering → judgment call; lean toward a minimal glossary covering just proper nouns and the 3-5 most repeated terms.

A glossary is not a bureaucratic ritual. It exists to prevent the same concept from being rendered three different ways across a long text. If the text has nothing to drift on, skip it.

Always extract:

• Proper nouns — people, places, organizations, products, publications. Decide once on transliteration vs. translation vs. preserving original script. Note whether to give the full name on first mention and a short form afterward.
• Technical and domain terms — pick the target-language equivalent and use it consistently.
• Recurring phrases or leitmotifs — if the author deliberately repeats a phrase (a coinage, a refrain, a key concept), preserve the repetition with a single fixed rendering.
• Numbers, dates, measurements — decide on format (1,000 vs 1.000; 5 km vs 3.1 mi; ISO dates vs locale dates) and whether to convert units. Be consistent.
• Acronyms — spell out on first mention with the original in parentheses, or use the target-language standard form if one exists.
• Titles of works — use the official translated title if one exists; otherwise transliterate or translate with the original in parentheses on first mention.
• Quotations from third sources — locate the canonical published translation if it exists; otherwise translate faithfully and flag the passage in the translator's notes.

Glossary format:

For long texts or texts with heavy domain terminology, show the glossary to the user before starting the translation and ask whether to proceed. For shorter texts (under ~1000 words), include the glossary inline at the top of the response and move on.

Phase 3 — Strategy decisions

State the policy — briefly to yourself, and to the user when asked — for:

• Domestication vs foreignization. How much to adapt cultural references. Default to a middle stance: preserve names and culturally specific concepts, but unpack opaque references with a brief in-text clarification or a footnote.
• Idioms and set expressions. Apply this priority order:
  1. If the target language has a natural equivalent idiom → use it.
  2. If no equivalent but the figurative meaning carries clearly → paraphrase plainly. Do not invent a clunky calque just to mirror the original's figurativeness.
  3. If the expression is culturally or stylistically load-bearing (a famous quote, an allusion the author is clearly building meaning from, a phrase that is itself the point) → keep it (transliterated or rendered literally) and add a footnote explaining it.
  4. Never invent a target-language idiom that does not actually exist just to "match" the source's idiomatic register.
• Units, dates, currencies. Convert (e.g., miles → km, °F → °C) only when the original is communicating a sense of magnitude that the target audience would otherwise miss. Otherwise preserve and let context carry.
• Punctuation and typography. Switch to target-language conventions: en-dash and em-dash usage, «guillemets» vs "quotes" vs „lower-upper", spaces before punctuation (French), non-breaking spaces, ellipsis style, etc.
• Paragraph and section structure. Preserve. Do not merge or split paragraphs except to avoid genuinely unidiomatic results.

Phase 4 — Translate in coherent units

Translate in paragraph-sized or larger semantic units, never sentence-by-sentence. Sentence-atomic translation preserves source syntax at the cost of argumentative flow, repetition, and cohesion — it is the second most common quality defect.

For each paragraph (or natural semantic unit):

1. Read the whole unit in the source.
2. Identify its function in the larger argument: claim, evidence, transition, aside, punchline, scene-setting.
3. Render it in the target language as a native writer would, preserving that function and the rhythm.
4. Apply the glossary consistently for every term.
5. Check the seam with the previous paragraph — does the cohesion connect cleanly (anaphora, "however", "by contrast", pronoun references)? Adjust if not.

Sentence boundaries can shift. Two short source sentences may merge in the target; one long source sentence may split. Do this only when it produces a more natural result, not as a default move.

Style preservation tactics:

• Match sentence-length distribution where possible — don't smooth a deliberately jagged style into uniform medium-length sentences.
• Preserve fragments, one-word paragraphs, em-dashes, parentheticals — these are usually deliberate authorial moves.
• Match density of clauses, not just word count.
• If the source uses repetition for emphasis, keep the repetition. Synonym variation that the source did not have is an addition.
• Resist the urge to "improve" the original. The author's quirks are the style.

Phase 5 — Footnotes

Use translator footnotes sparingly. Each footnote interrupts reading, so the bar is high. Use them only when one of these holds:

• A source expression was preserved (rather than adapted) and the target reader cannot decode it without help.
• A cultural reference is load-bearing for the argument and not self-explanatory in the target context.
• A genuine ambiguity in the source had to be resolved in one direction, and the user might want to know.

Format:

• Mark in text with a superscript number or [1] after the relevant phrase.
• Collect notes at the end of the translation under a heading like **Translator's notes** or its target-language equivalent.
• Each note is one or two sentences. No mini-essays.

Do not footnote things the target reader can easily look up, or things with an obvious target-language equivalent — that is avoidance of the actual translation work, not service to the reader. Aim for fewer than ~1% of words requiring footnotes. If more, rethink the cultural adaptation strategy.

Never silently drop a difficult passage or paraphrase it past recognition. If a passage is genuinely ambiguous, resolve it one way and footnote the ambiguity.

Phase 6 — Self-review before delivering

Run an explicit pass against the criteria. This is not optional — it is where most defects are caught.

• Drift check. For each glossary entry, confirm the chosen rendering was used every time the source term appears. Search the target text for unintended variants. Drift typically starts around the 60–70% mark of a long text — check there first.
• Omission check. Walk through the source paragraph by paragraph; confirm every proposition is present in the target.
• Addition check. Did any clarification or interpretation get smuggled in that the source did not have? If yes, either remove it or move it to a footnote.
• Naturalness check. Mentally read the target aloud. Any sentence that sounds translated should be rewritten.
• Register check. Sample 3–4 sentences from different parts of the text — do they all sit in the same register?
• Style check. Pick one paragraph from source and target. Same rhythm? Same density of clauses? Same use of dashes, parentheticals, fragments?

Output structure

Default output, in this order:

1. Glossary — for texts over ~800 words, or shorter texts with dense domain terminology. Mark the section so the user can skip or collapse it.
2. Translation — full text in the target language, mirroring the source's structure (headings, paragraphs, lists, blockquotes, code blocks all preserved).
3. Translator's notes — only if footnotes were used or non-obvious decisions were made.
4. Decision log — optional, 2–4 lines, only when significant judgment calls were made that the user might want to know about. Examples: "Converted miles to km because the article relies on a sense of distance"; "Kept «гласность» untranslated with a footnote because glasnost carries historical weight that 'openness' loses".
5. Flags for review — optional, for long or technically dense texts. A short list of specific passages where the source was ambiguous and you had to choose an interpretation, terms where you were unsure of the target-language convention, or cultural references whose adaptation you'd want a second opinion on. Cite the passage by quoting 3–5 source words and stating the choice. Skip this section if there is nothing genuine to flag — do not manufacture flags for symmetry.

For shorter texts, drop sections 1, 4, and 5 unless they would actually help the user.

File output

By default, save the translation as a separate file next to the original. Override only when the user explicitly asks for inline output.

Source is an uploaded or referenced file (e.g., /path/to/article.md):

• Write the translation to a new file in the same directory as the original, mirroring the original's extension where possible.
• Name it <original-basename>-<target-lang>.<ext>, using a two-letter language code (en, ru, de, fr, es, pt, zh, ja, etc.). Example: article.md → article-en.md. If the original already carries a language code (article-ru.md), swap it (article-en.md) rather than appending.
• Place the glossary, footnotes, and translator's notes inside the same file, in the order described in "Output structure" above. Do not split into multiple files unless the user asks.
• In the chat reply, point the user to the new file path and surface what needs attention — the glossary if substantial, decisions worth flagging, ambiguities resolved — without pasting the full translation back into chat.
• For complex source formats (.docx, .pdf, .pptx), use the corresponding document-creation skill, or ask the user whether a plain .md output is acceptable instead.

Source was pasted directly into the chat (no file path to anchor to):

• Produce the translation inline in the structure described in "Output structure".
• If the result is long (>~2000 words), offer to save it to a file — but do not save automatically without asking, since there is no original location to put it next to.

Inline output is correct (not the file default) when:

• The user explicitly asks for it ("just paste it here", "show it in chat", "in the response").
• The text is very short (under ~200 words) and a file would be overkill.
• The user is iterating on a specific passage or sentence rather than translating a whole document.

Formatting and special cases

Mechanical formatting rules (headings, links, code, tables, YAML front matter, HTML embeds) and special-case playbooks (embedded quotations, mixed-language sources, headlines and pull quotes, very long texts >5000 words) live in references/formatting.md. Read it before Phase 4 whenever the source is anything richer than plain paragraphs. The one rule that never bends: when in doubt about a syntactic element, leave it verbatim — breaking the document's rendering is worse than under-translating an attribute value.

Final note

Every translation choice should be defensible against the criteria above. When in doubt, prefer the rendering that a thoughtful native reader of the target language — one who also knows the source language — would call "obviously the right way to put it".