Read As Markdown

Read as Markdown

Convert PDF, DOCX, and EPUB files to cached Markdown. The converted Markdown is stored in a per-workspace cache so repeated reads are instant.

Quick start

uv run ~/.claude/skills/read-as-markdown/scripts/read_as_markdown.py ~/documents/paper.pdf

Returns JSON with the cache path, total line count, and the first 200 lines of content. Use the Read tool with offset/limit on cache_path for further chunks.

How it works

1. Detects format from file extension
2. Finds workspace root (walks up to find .claude/ directory)
3. Checks cache at <workspace>/.cache/markdown/<relative-path>.md
   • Fast path: mtime + size match in .meta sidecar → cache hit
   • Content-hash fallback: SHA-256 of original file checked against index.json → finds cache even if the source file was moved/renamed
4. On miss: dispatches to the appropriate converter backend
5. Writes markdown + .meta sidecar + updates index.json hash index
6. Returns JSON with first N lines (default 200)

CLI interface

read_as_markdown.py FILE [options]

positional arguments:
  FILE                    Path to document file (.pdf, .docx, .epub)

options:
  --workspace-root PATH   Override workspace root (auto-detected via .claude/)
  --limit N               Max lines to return (default: 200)
  --offset N              Line offset for chunked reading (default: 0)
  --backend NAME          Force a specific backend (pandoc, pdftotext)

Output format

{
  "cache_path": "/home/user/.cache/markdown/downloads/paper.md",
  "total_lines": 847,
  "content": "first 200 lines...",
  "source": "/home/user/documents/paper.pdf",
  "format": "pdf",
  "backend": "pdftotext",
  "cached": false
}

On error: {"error": "description"}.

Caching

Cache lives at <workspace>/.cache/markdown/ mirroring the source file tree. Each cached file has a .meta JSON sidecar:

{
  "mtime": 1709312456.789,
  "size": 2048576,
  "content_hash": "sha256:a1b2c3...",
  "toolchain": [
    {"step": "convert", "backend": "pdftotext", "version": "24.04.0", "at": "2026-03-01T12:00:00+00:00"}
  ],
  "converted_at": "2026-03-01T12:00:00+00:00"
}

A hash index at <workspace>/.cache/markdown/index.json maps content hashes to cache paths, enabling cache hits when source files are moved or renamed.

The toolchain array records every processing step — the initial conversion and any later post-processing (formatting, cleanup, refinement by helper scripts). Each step records its name, tool, version, and timestamp.

Backends

pdftotext uses -layout mode for best table/column preservation in academic papers. On NixOS, automatically falls back to nix-shell -p poppler-utils if pdftotext is not in PATH.

pandoc uses --to=markdown --wrap=none for clean, unwrapped output.

Adding a new backend

Add a converter function and register it in BACKENDS:

def marker_converter(source: Path) -> tuple[str, str]:
    # ... convert ...
    return markdown_text, "marker 1.0"

BACKENDS["pdf"].insert(0, ("marker", marker_converter))

Distinction from pdf/docx skills

• read-as-markdown: Reads documents → produces Markdown for the agent to consume
• pdf skill: Creates, merges, splits, watermarks, OCR, fills forms in PDFs
• docx skill: Creates, edits, tracks changes in Word documents