Multimodal Source to Markdown Converter

Problem

This skill addresses the case where raw material cannot be fed directly into later analysis: notes, recordings, screenshots, scanned pages, spreadsheet exports, saved web docs, and ZIP packages are mixed together, and LLM workflows can easily drift into summaries, classifications, or recommendations. source-to-markdown only converts source files into text and does not perform requirement analysis, product scoping, hypothesis extraction, recommendations, or handoff generation, making it suitable when the goal is to preserve the original evidence as closely as possible.

How It Works

It routes by file type: md and txt are read directly; docx, pdf, pptx, xlsx, csv, html, json, and xml go through MarkItDown; images use Tencent OCR by default with a visual LLM fallback for layout-dependent transcription; audio uses Tencent ASR by default with an LLM audio fallback when needed; zip packages are expanded and produce manifest.md. Outputs use .md for documents and structured exports, and .raw.md for OCR, ASR, and LLM transcription. Unclear content is marked as [unrecognized] or [unclear], and missing text is not invented.

Boundaries

Video, YouTube URLs, EPUB, and arbitrary binary files are not supported by default. Failures are recorded with the error message and a simpler source may be requested. API credentials in providers.json must not be pasted into chat, logs, or converted output. The intended result is verifiable raw text, not analytical conclusions.