Translating a document so the formatting stays exactly the same

First working version
< 1 h

Goal

A skill that translates the document by meaning while leaving names, figures, bank details and URLs alone, lifts out the logos, images and stamps to put them back at the same point on the page, watches that the translation does not overflow a single cell or text box and checks that what it hands over is still the same document with another language inside it.

How it works

First it asks for the target language, as clickable options, because that is the one thing it cannot work out from the file. The source language it detects on its own, and it only asks about register when the document does not make it obvious.

Before translating anything it walks the document and writes down what is there: structure, styles, images and everything that must not be touched. The inventory is taken first for a specific reason, which is that it later becomes the list everything is verified against; drawn up at the end it checks nothing. The long part of that list is what does not get translated: names of people and companies, brands, text inside a logo, code, URLs, email addresses, figures, dates, bank and tax identifiers and product codes.

Then it translates by meaning rather than word by word, which is the difference between a document you can use and one that has to be redone: it uses the established equivalent in technical and legal terms, and when it hesitates between translating and leaving the original it leaves the original and adds the translation in brackets the first time. All the while it watches for overflow, because on a slide or in a cell longer text does not get cut, it breaks the box.

The document is never rebuilt: inside the file only the text is replaced, and where the page has to be rewritten the logos, images, signatures and stamps are extracted with their position, size and layer order and put back at the same coordinates, without rescaling. Text embedded in a logo is not translated and the image is not regenerated: if it matters, its translation goes in the report.

To finish, it goes back through the inventory point by point: no text left untranslated, page, table and cell counts still reconciling, images and signatures still at the same coordinates and the same size, and the data that does not get translated identical character by character.

Agent steps
Deliverable
  • A new copy with the target language code in its name, in the same format as the original.
  • A report of which terms were left untranslated and why.
  • The result of every verification check and the points worth looking at yourself.
Stack
  • Claude Cowork
  • A Claude skill of our own
  • Word, PowerPoint and Excel skills
  • PDF skill
  • Up-front inventory

Why it matters

Translating a working document is not just translating its text: it has to end up looking exactly the same. The value is in the tables still adding up, the logo and the stamp still sitting at the same point on the page and the account number having stayed exactly as it was, and that takes understanding what each thing inside the document is: telling a clause, which gets translated, from an IBAN or a product code, which do not, and lifting the graphic elements out to put them back where they were instead of rebuilding the page around them.

What could come next

  1. 01A per-client glossary, so the same technical term is always translated the same way across documents and across jobs. Today each translation decides on its own, and each decision is reasonable in isolation, but two documents for the same client can end up calling the same part two different things.
  2. 02Translate into several languages in one job and hand over the whole folder. Right now the work is repeated language by language, inventory and verification included, and those two are exactly the same all four times.
  3. 03Flag the legal or technical passages that deserve a human pass, instead of leaving a generic warning at the end. A warning that covers the whole document gets read once and forgotten; pointing at the three paragraphs that actually matter is what makes someone look at them.
  4. 04Update the ERP with the translated version hanging off the same client record as the original, so whoever handles a foreign account does not have to ask by email which document is the good one.

Worth knowing

  • legal

    A translated contract is not a valid contract in the other language. Anything to be signed or filed with an authority still needs a sworn translation, and no automation replaces that.

  • dependency

    In .docx, .pptx and .xlsx fidelity is exact because only the text is swapped. In a scanned PDF there is no layer to translate: what can be offered is a rebuilt .docx, a faithful approximation but not a visual clone.

  • dependency

    Overflow is the usual failure and it does not always show up in the text: a slide in German runs longer than in Spanish and the box breaks. It gets adjusted, but decks are worth a look before they go out.

The prompt

To use it, just copy this prompt and paste it into Claude.

Create a new, self-contained SKILL for translating documents while preserving the original formatting exactly. The skill translates the document into the target language the user specifies, produces a new translated copy and never touches the source file.

STEP 0 (before generating anything): ask me by click, multi-select, which formats the skill must support. Offer these options: Word (.docx), PowerPoint (.pptx), Excel (.xlsx), PDF (.pdf). Generate support only for the ones I select. If I select several, the skill must handle them all. If I select none, ask me again.

CONTEXT
A skill for ForgeLab, professional deliverable standard. The translated document must look identical to the original: layout, images, logos, typography, styles, tables and charts. It will be used regularly on client documents, so fidelity and verification are what matter most.

FORMATS
Fidelity is exact in .docx, .pptx and .xlsx, because only the text is replaced. In .pdf with a text layer it is high; in a scanned .pdf, or one with no reliable layer, the skill flags the limitation and offers an alternative before continuing.

MATERIALS (read them before generating anything, only for the selected formats)
- /mnt/skills/public/docx/SKILL.md
- /mnt/skills/public/pptx/SKILL.md
- /mnt/skills/public/xlsx/SKILL.md
- /mnt/skills/public/pdf/SKILL.md
They hold the environment's constraints (libraries, paths, rendering). The SKILL you generate must state that, before touching each document, the public SKILL for its format is to be read.

RULES FOR THE SKILL YOU ARE GOING TO GENERATE
1. Always a new copy. Never edit or overwrite the source. Name the copy with the target language code (example: contract.docx becomes contract_EN.docx) and keep the same format as the original.
2. Ask only what is essential, by click. Before translating, present clickable options (with free input available) for target language (compulsory, never assumed) and register (formal, technical, commercial, neutral) only if it cannot be inferred from the document. The source language is detected and only confirmed if there is real doubt.
3. Offer the format to the user. If it is unclear what is being translated, show the supported formats as clickable options to choose from.
4. Take inventory before translating. Walk the document and note structure, styles, images and everything that must not be translated. That is the verification checklist.
5. Exact formatting: layout, breaks, tables, columns, numbering, bullets, headers, footers and notes.
6. Do not alter graphic elements: images, logos, charts, signatures or stamps. The document is not rebuilt: in .docx, .pptx and .xlsx only the text is replaced and no graphic part of the file is touched. When the page has to be rewritten, extract every image, logo, signature and stamp with its position, size, rotation and layer order, and put them back at exactly the same coordinates, without rescaling or recompressing. Text embedded in an image is not translated and the image is not regenerated: if it matters, the translation of that text goes in the report.
7. Do not translate: proper names of people and companies, brands, text embedded in logos, code, URLs, email addresses, figures, dates, bank account and tax identifiers, addresses and product codes.
8. Translate by meaning, not word by word. Use the established equivalent for technical or legal terms. When torn between translating and leaving the original, leave the original and add the translation in brackets the first time.
9. Do not add, summarise or remove content. Translate all translatable text, including text in tables, notes, captions and boxes.
10. Control overflow. If the translation takes up more room, adjust so that cells, text boxes and shapes do not overflow (critical in .pptx and .xlsx).
11. The PDF-without-text-layer case. If the PDF is scanned or has no reliable layer, warn with clickable options and propose delivering it as a rebuilt .docx, making clear it will be a faithful approximation and not an exact visual clone. Wait for a decision.
12. Verify before delivering, point by point: no text is left untranslated; page, paragraph, table and cell counts match; every image, logo and signature is still at the same coordinates and the same size; styles untouched; nothing overflows; and the untranslatable data is identical character by character. If anything fails, fix it and verify again.
13. Be honest about quality. Textual and structural verification is reliable; the fine visual check is not infallible. Do not promise guaranteed perfection: deliver verified high quality and point out what is worth reviewing.

WHAT THE SKILL DELIVERS IN USE
When used, the skill delivers the translated document plus a short report with: terms left untranslated and why, a verification checklist with the result of each point, and the points the user should review personally.

OUTPUT FORMAT FOR THIS REQUEST
Generate a single SKILL.md at /mnt/skills/user/[SKILL_NAME]/SKILL.md with:
- YAML frontmatter with name and description. The description must trigger the skill well: include only the formats selected in STEP 0 and trigger phrases in Spanish and English (traduce este documento, traduce este Word/PDF/PowerPoint/Excel, traduce manteniendo el formato, translate this document keeping the format).
- Body in Spanish, lean register (no filler, no decorative subheadings), with the workflow as numbered steps and the rules section, limited to the selected formats.
Ask me for the skill's name before creating it if I have not given it to you. Do not edit any other file.

Contact

Let's have a call. In 30 minutes I can tell you whether I can help.

Email address (enable JavaScript to reveal)LinkedIn