Processing WhatsApp Exports

Converts raw WhatsApp exports into clean, chunked markdown files ready for reading. Fully offline.

Workflow

1. Process — run process-conversation.py. Safe to re-run: it extracts into exported/, regenerates all chunks. A newer export just overwrites and regenerates.
2. Check output — the script prints a summary including pending audio transcriptions.
3. Ask about transcription — if there are pending audio messages, ask the user if they want them transcribed.
4. Transcribe (if yes) — run transcribe-conversation-audios.py (see below).
5. Re-process — run process-conversation.py again to regenerate chunks with transcriptions inline.
6. Report — tell the user the output location and summary.

Running the processor

python3 ${CLAUDE_SKILL_DIR}/scripts/process-conversation.py "path/to/export.zip" --output-dir .

Always pass --output-dir . to create the conversation folder in the current project directory.

Options: --output-dir DIR (default: same parent as source), --max-chunk-kb N (default: 100).

Re-running with a newer export of the same conversation overwrites exported/ and regenerates all chunks. Existing transcriptions (.txt files) are preserved if the audio files haven't changed.

Transcribing Voice Messages

python3 ${CLAUDE_SKILL_DIR}/scripts/transcribe-conversation-audios.py "path/to/whatsapp-name-conversation" <language>

Transcribes all .opus files in exported/ that don't have a .txt yet. Saves the language and transcription preference in .processed. After transcription, re-run process-conversation.py to regenerate chunks with inline transcriptions. On future re-runs, process-conversation.py auto-transcribes new audio files if .processed has transcribe: true.

Language detection: Read the first ~50 lines of the conversation to identify the language (<attached: = en, <anexado: = pt, <archivo adjunto: = es). Pass as whisper code: en, pt, es, etc.

Output structure

whatsapp-name-conversation/
  .gitignore              # ignores exported/
  .processed              # metadata (transcription preference, language)
  chat-2024-01-01_2024-06-15.md
  chat-2024-06-16_2024-12-31.md
  exported/               # gitignored — full export contents
    _chat.txt
    00000001-PHOTO-*.jpg
    00000002-AUDIO-*.opus
    00000002-AUDIO-*.txt  # transcriptions live alongside audio

Markdown format

**John:**
[08/04/2026, 14:30:00] Hello there
[08/04/2026, 14:31:00] [Photo](exported/00000001-PHOTO-abc.jpg)
[08/04/2026, 14:32:00] 🎤 "Transcribed voice message text here"
> Audio: [exported/00000003-AUDIO-abc.opus](exported/00000003-AUDIO-abc.opus)
[08/04/2026, 14:33:00] [Document: contract.pdf](exported/00000005-DOC-contract.pdf)
[08/04/2026, 14:34:00] [Voice message](exported/00000004-AUDIO-def.opus)

**Jane:**
[08/04/2026, 14:35:00] Got it, thanks!

Messages are grouped by sender — the name only appears when the sender changes. Non-transcribed voice messages appear as [Voice message](exported/AUDIO-xxx.opus).

Reading processed conversations

• Chunk files are ~100KB each — read them sequentially or search with Grep
• ls *.md in the output folder shows date ranges at a glance
• Images: Read tool (multimodal — reads directly from exported/)
• PDFs: Read tool (native PDF support from exported/)
• Videos: Cannot be read — note and move on
• Audio: Read the transcription inline in the markdown, or read the .txt file in exported/

Dependency Check

${CLAUDE_SKILL_DIR}/scripts/check-deps.sh

Run this before processing to verify all tools are available.

Whisper Setup

Transcription uses whisper.cpp locally. Fully offline — no API keys needed.

First-time setup (requires ffmpeg and cmake — install with brew install ffmpeg cmake):

Check if whisper.cpp is already installed (~/.whisper.cpp/build/bin/whisper-cli or locally in the skill directory). If not found, ask the user whether to install shared (~/.whisper.cpp, reusable across projects) or local (inside the skill directory). Then run:

${CLAUDE_SKILL_DIR}/scripts/setup.sh --shared   # installs to ~/.whisper.cpp
${CLAUDE_SKILL_DIR}/scripts/setup.sh --local     # installs to skill directory

This clones whisper.cpp, builds it, and downloads the large-v3 model (~3GB). Guide the user through installation if prerequisites are missing.

Performance: ~2-10 seconds per voice message on Apple Silicon.

Localization

The parser handles multiple languages for attachment markers:

• EN: <attached: filename>
• PT-BR: <anexado: filename>
• ES: <archivo adjunto: filename>

Tips

• Payment receipts are usually photos (PHOTO-) sent right after messages mentioning amounts
• Photos can be documents — screenshots, photos of paper, receipts
• Financial information often appears in message text, not just in attachments
• Filenames may have corrupted accented characters — this is normal