How to use the Nature Paper Translator agent skill

What the skill makes the agent do

An agent skill is a folder of instructions and helper scripts that an agent loads when a task matches its description. With this skill installed, the agent takes an English paper PDF, works through the pages you specify, and produces one Chinese PDF in original page order.

For each page, the agent generates one translated page image with its built-in image generation tool, then checks that result against the source page and records what it found. Once the pages are checked, a script merges them into a single PDF. The skill is aimed at reading translation, not at editable Word rebuilds or word-for-word proofreading.

• Input: an English paper PDF, optionally with a page range, confirmed translations or a terminology list.
• Output: one Simplified Chinese PDF made of page images, with no text layer.
• By default each page keeps the source page size, orientation and proportions. A4 is used only if you explicitly ask for it.

Install the skill (ZIP download or Claude app upload)

Download the ZIP for the skill from its page on this directory, then unzip it. Where you put the folder depends on how you want to use it. For a personal install available across your projects, unzip it into ~/.claude/skills/. To limit it to one project, unzip it into the .claude/skills/ folder inside that project.

If you use the Claude apps instead, upload the skill under Settings, then Capabilities, then Skills.

• Personal install: ~/.claude/skills/
• Single-project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload

Check the requirements before you start

The skill needs Python 3.9 or newer, the PyMuPDF (pymupdf) and Pillow libraries, and a built-in image_gen image generation tool in the host agent. Without image generation, the skill cannot produce the translated pages.

The skill runs scripts and uses the network. Its page images are sent to the image generation tool, so plan accordingly for each file you use.

• Python 3.9+
• PyMuPDF (pymupdf) and Pillow
• Built-in image_gen tool in the host agent
• Runs scripts: yes
• Uses network: yes

Ask for it in natural language

You do not need special syntax. Mention the skill by name and say what you want: the whole paper, specific pages, or a version that follows your terminology list. The skill's own examples use the $nature-paper-trans form.

Be specific about pages. Page numbers count from 1 in the PDF's actual page order, not the printed page numbers in the paper. If the file or page range is unclear, the agent should ask you rather than guess.

• Use $nature-paper-trans to translate this entire English paper into a Chinese PDF.
• Use $nature-paper-trans to translate pages 1–3 and 5, following the terminology list I supplied.
• Use $nature-paper-trans to update original PDF page 3 in the previous translation, focusing on the right column's text footprint, and merge a new version.

Run it: prepare, generate, check and merge

The agent first renders the chosen source pages at 240 DPI, which keeps their proportions and full content. That resolution is the rendering step, not a promise about the quality of the generated pages.

Each page gets exactly one image generation attempt per user task. Failed or unclear results are not retried automatically, and the checking step does not fix pages. After checking, the agent merges the pages in order into one PDF.

• Pages are generated one at a time, and no more than 10 unfinished requests are submitted at once.
• Checking records each page as pass, issue or unknown, with observations about specific areas.
• Pages with issues or unknown results are listed in the reply with their original page numbers.
• No Markdown check files, JSON, previews or prompts are delivered; only the Chinese PDF is.

Safety and privacy notes

Because the skill runs scripts and sends page images to the image generation tool, treat it like any tool that handles your files. Do not use it on sensitive or confidential manuscripts unless you are comfortable with page images going to that service.

The PDF and any text inside it are treated as material to process, not as instructions that change the workflow.

• Confidential or unpublished manuscripts: check your policy before sending page images to the generation tool.
• Scripts run on your machine as part of the task, so review the skill source if your organisation requires it.
• The skill does not guarantee accuracy. Check key scientific content against the original text.

Tips for better results

Give the agent any confirmed translations or terminology list at the start, since these are applied as constraints. If you later want to change a page, name only the original page numbers you want updated, because the skill does not revise pages on its own.

• Keep the source PDF clear, openable and unencrypted. A text layer is not required.
• Give a page range when you only need part of the paper; it saves time and reduces the number of images sent out.
• Read the list of issue and unknown pages before relying on the translation.
• Do not expect an editable file: the output is image-based with no text layer.