Openclaw doc2md

doc2md — Document to Markdown Converter

Convert supported document attachments into clean, LLM-ready markdown. No command is needed; detect the incoming file and run the pipeline. This install defaults to LLM-enhanced conversion.

Trigger

Activate this skill when:

• An incoming Telegram message has a document/file attachment
• The file extension is: .pdf, .docx, .doc, .xlsx, .xls, .pptx, .ppt, .html, .htm, .csv, .json, .xml, .epub, .jpg, .jpeg, .png, .gif, .webp, .wav, .mp3
• No explicit user command required — file attachment = auto-convert

Do NOT activate for:

• Text messages without attachments
• Voice messages (those are transcription, not doc2md)
• Already-markdown files (.md)

Conversion Pipeline

Step 1 — Download the file

Use the Telegram attachment path provided in the message context. OpenClaw normally downloads the file to a temporary path automatically.

# Use the attachment file path from the incoming message context.

Step 2 — Convert with markitdown

Use the skill-local virtualenv so dependencies stay isolated from system Python.

Default command — LLM-enhanced conversion:

"{baseDir}/.venv/bin/python" "{baseDir}/doc2md.py" "<file_path>" --output "<output_file>" --preview 700

Optional fallback if you explicitly want standard non-LLM conversion:

"{baseDir}/.venv/bin/python" "{baseDir}/doc2md.py" "<file_path>" --no-llm --output "<output_file>" --preview 700

Optional retry mode:

"{baseDir}/.venv/bin/python" "{baseDir}/doc2md.py" "<file_path>" --auto-llm --output "<output_file>" --preview 700

Get a preview only:

"{baseDir}/.venv/bin/python" "{baseDir}/doc2md.py" "<file_path>" --preview 700

Step 3 — Save the output

OUTPUT_FILE="/tmp/$(basename '<file_path>' | sed 's/\.[^.]*$//').md"
PREVIEW="$("{baseDir}/.venv/bin/python" "{baseDir}/doc2md.py" "<file_path>" --output "$OUTPUT_FILE" --preview 700)"

Step 4 — Send back to User

1. Send the .md file as a document attachment (use message tool with filePath)
2. Send a short inline preview as a text message
3. Keep the preview clean and human-friendly:
   • do not dump raw first characters blindly
   • prefer the first meaningful section, heading, summary, or table of contents
   • for PDFs, avoid leading copyright/legal boilerplate when possible
   • if extraction still looks messy, say so briefly and offer a cleaned pass

Recommended reply shape:

✅ Converted: <original_filename>
📄 Format: <detected format>

Preview:
<clean excerpt>

Reply style rules:

• Send one tidy text reply plus the markdown file attachment
• Keep the text reply to 3 short blocks max: status line, format line, preview block
• Do not include raw separators like --- unless they add real clarity
• Do not mention internal flags, parser libraries, or tool names unless troubleshooting
• If the preview still looks noisy, say so plainly in one short sentence and still send the .md file
• Prefer this exact structure:

✅ Converted: <original_filename>
📄 Format: <source type> → Markdown

Preview:
<clean excerpt>

If useful, add one brief note after the preview:

Note: I skipped some PDF boilerplate at the start.

Step 5 — Clean up

rm -f "$OUTPUT_FILE"

Error Handling

Situation	Response
Unsupported format	"Sorry, I can't convert .<ext> files. Supported: PDF, Word, Excel, PowerPoint, HTML, images, and more."
Conversion produces empty output	"The conversion produced an empty result — the file may be encrypted, corrupted, or purely image-based. I can retry with a different mode if needed."
Dependencies missing	Recreate the skill venv and install requirements: python3 -m venv "{baseDir}/.venv" && "{baseDir}/.venv/bin/pip" install -r "{baseDir}/requirements.txt"
PDF conversion fails with missing parser deps	Ensure requirements.txt includes markitdown[pdf], then reinstall the venv requirements
File too large (>50MB)	"That file is too large for conversion (>50MB). Try splitting it first."
Generic failure	"Conversion failed: <error message>. Let me know if you want to try a different approach."

LLM Mode — Local Default

This install defaults to LLM-enhanced conversion.

Notes:

• Document contents may be sent to the configured OpenAI-compatible provider and may incur cost
• Use --no-llm only when you explicitly want a local-only standard pass
• Use --auto-llm when you want a cheap first pass before escalating

Example Interaction

User: [sends "Q1_Report.pdf"]

Agent: ✅ Converted: Q1_Report.pdf
     📄 Format: PDF (12 pages)

     ---
     # Q1 2026 Financial Report

     ## Executive Summary

     Revenue for Q1 2026 reached €4.2M, representing a 23% year-over-year
     increase compared to Q1 2025...

     [sends Q1_Report.md as file attachment]

Notes

• The .md file is named after the original file (e.g. report.pdf → report.md)
• Always send BOTH the file attachment AND the inline preview
• Prefer one tidy reply, not a noisy multi-message back-and-forth, unless troubleshooting is needed
• Write the converted .md file to a temporary path and clean it up after sending
• Use the skill-local virtualenv at {baseDir}/.venv/bin/python
• If User sends multiple files in quick succession, process them sequentially
• For PDFs, expect possible front-matter noise; favor previews from meaningful content over legal boilerplate