PowerPoint Translation Skill

Translate PowerPoint presentations while preserving all formatting including fonts, colors, spacing, tables, and alignment.

When to Use This Skill

• Translating .pptx files between languages
• Batch translating multiple presentations in a directory
• Preserving slide formatting during translation (especially CJK ↔ English)
• QA-ing a translated deck with a vision language model (--vision-audit) to catch text overflow, truncation, garbled glyphs, untranslated leftovers, overlaps, or contrast problems
• When you need to inspect intermediate XML for debugging

Setup

Before first use, set up the environment:

# Navigate to the scripts directory
cd .claude/skills/ppt-translator/scripts

# Create virtual environment and install dependencies
python3 -m venv .venv
source .venv/bin/activate  # macOS/Linux
pip install -r requirements.txt

# Configure API keys
cp example.env .env
# Edit .env with your provider API key(s)

Basic Usage

cd .claude/skills/ppt-translator/scripts
source .venv/bin/activate

python main.py /path/to/presentation.pptx \
  --source-lang zh \
  --target-lang en

(The default provider is DeepSeek deepseek-v4-flash.)

Provider Configuration

CLI Reference

Output Files

For each input presentation.pptx, the tool generates:

1. presentation_original.xml - Extracted source content (deleted unless --keep-intermediate)
2. presentation_translated.xml - Translated content (deleted unless --keep-intermediate)
3. presentation_translated.pptx - Final translated presentation with formatting intact
4. presentationtranslatedvision_audit.md - Only with --vision-audit: per-slide findings from the vision model, grouped by slide with severity (high/medium/low)

Common Workflows

Translate a Single File (Chinese → English)

python main.py deck.pptx --provider anthropic --source-lang zh --target-lang en

Batch Translate a Directory

python main.py /path/to/presentations/ --provider openai --source-lang ja --target-lang en

Debug Translation Issues

python main.py deck.pptx --keep-intermediate --provider deepseek
# Inspect the generated XML files to see extracted/translated content

Use a Specific Model

python main.py deck.pptx --provider openai --model gpt-5-mini

Translate with Gemini (Cost-Effective)

python main.py deck.pptx --provider gemini --source-lang ko --target-lang en

Translate and Run the Visual Audit

python main.py deck.pptx --source-lang zh --target-lang en --vision-audit

After rebuilding the deck, every slide is rendered to a PNG (LibreOffice headless → pdftoppm) and sent to deepseek-v4-flash-vision-exp together with the slide's expected translated text. The model compares the rendering against the expected content and reports concrete visual defects. Read {deck}translatedvision_audit.md afterwards and fix any HIGH-severity findings (e.g. by shortening text or enlarging boxes) before shipping the deck.

Audit-only reruns on an already-translated deck are not needed — the audit runs automatically at the end of a --vision-audit run, and a failed audit never aborts the translation itself.

Supported Languages

Use standard ISO 639-1 language codes:

Design Notes

Font Scaling

The tool automatically scales fonts down (70% for text, 80% for tables) to accommodate text expansion when translating from compact languages (Chinese, Japanese, Korean) to English. This prevents text overflow in fixed-size text boxes.

Caching

Repeated strings within a presentation are cached to avoid redundant API calls. This is especially useful for presentations with recurring headers, footers, or terminology.

Chunking

Long text blocks are intelligently split at sentence boundaries to stay within API limits while preserving translation quality.

Troubleshooting

"API key not found"

Ensure your .env file in the scripts directory contains the correct environment variable:

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

Formatting looks wrong

1. Use --keep-intermediate to inspect the XML files
2. Check if the source presentation has unusual formatting
3. Try a different provider

Translation incomplete

1. Check for API rate limits with your provider
2. Try reducing --max-chunk-size for very long text blocks
3. Ensure your API key has sufficient quota

"Visual audit skipped: ... tools that were not found"

The visual audit needs LibreOffice (soffice) and poppler (pdftoppm):

brew install --cask libreoffice && brew install poppler   # macOS
# apt install libreoffice poppler-utils                   # Debian/Ubuntu

If LibreOffice itself crashes on launch (seen on some macOS 26 + LibreOffice 25.x combos), upgrade it: brew upgrade --cask libreoffice.

"The visual audit requires a DeepSeek API key"

The audit always calls DeepSeek's vision endpoint regardless of the translation provider, so DEEPSEEKAPIKEY must be set even when translating with another provider.

Script Reference

The scripts/ directory contains:

• main.py - Entry point
• requirements.txt - Python dependencies
• example.env - Environment variable template
• ppt_translator/ - Core translation module

- cli.py - CLI argument parsing - pipeline.py - PPT extraction and regeneration - translation.py - Chunking and caching - providers/ - LLM provider implementations - vision_audit.py - Slide rendering (LibreOffice + pdftoppm) and DeepSeek vision-model QA