MarkItDown

Convert a document, image, audio file, or YouTube URL to Markdown using Microsoft's markitdown CLI. The skill validates the input, composes the right flags, optionally saves the result under ~/.agents/output/<project>/markitdown/<slug>/, and reports a one-line summary with the fully-expanded absolute path (no tilde, no magic).

The deterministic work — install check, validation, slug derivation, save path, command composition — happens in scripts/markitdown.sh. The skill parses $ARGUMENTS, hands them to the script, and turns the script's RESULT: lines into a human report.

Install

pip install 'markitdown[all]'

For a smaller install, pick only what you need:

For Azure Document Intelligence, also export MARKITDOWNDOCINTELENDPOINT=https://<resource>.cognitiveservices.azure.com/ before invoking with -d.

Parameters

Output saved under ~/.agents/output/{project}/markitdown/{slug}/, where {project} is the kebab-cased basename of the git toplevel (else cwd) and {slug} is a kebab of the input basename (≤5 words). Pipeline-friendly — typical downstream: /forge -s -f <path> decomposes the extracted content into workstreams; /apex -f <path> implements from it; any skill accepting -f can consume.

Workflow

1. Resolve the target from arguments or unambiguous session context. Ask only when the target is missing or ambiguous. Before audio, OCR, Azure or plugin conversion, verify the installed backend and required data-transfer authorization; an optional dependency group does not imply local processing.
2. Run the helper:

$SKILLDIR = this skill's folder — ${CLAUDESKILL_DIR} in Claude Code, the directory containing this SKILL.md elsewhere.

``bash bash "$SKILL_DIR"/scripts/markitdown.sh -s '/absolute/path/report.pdf' ``

Parse arguments as data and pass each flag/value separately with shell quoting. Never splice raw $ARGUMENTS into shell code or use eval.

1. The script emits RESULT: key=value lines — keys: bytes, slug, saved, plus path when saving (order is not guaranteed; parse by key) — followed either by the converted Markdown (no-save mode, after a --- separator) or nothing (save mode — the file is on disk).
2. Parse the RESULT: lines and produce the report below.
3. If the script exits with ERR: markitdown not installed (exit 127) → print the install command from ## Install and stop. Never auto-install on the user's behalf.
4. If the script exits with another ERR: (file not found, missing endpoint, unknown flag) → relay the message verbatim and stop.

Output

markitdown: <input> → <bytes> bytes of Markdown
saved: <path>      # only when -s

When saving, report the actual file path. For long output, choose save mode when consistent with the request or capture stdout once in an authorized temporary file, then provide a bounded preview and the retained result. Do not repeat a costly conversion merely to truncate its display.

Examples

/markitdown ~/Downloads/report.pdf            # convert, print to terminal
/markitdown -s ~/Downloads/report.pdf         # convert + save under ~/.agents/output/<project>/markitdown/report/
/markitdown -s -p deck.pptx                   # use third-party plugins (e.g. markitdown-ocr)
/markitdown -d invoice.pdf                    # Azure Document Intelligence
/markitdown -k brand.html                     # keep base64 images inline
/markitdown https://youtu.be/dQw4w9WgXcQ      # YouTube transcript
/markitdown -l                                # list installed plugins, then exit

Notes

• YouTube URLs are detected by the https?:// prefix and passed straight to markitdown. The slug is derived from the URL's last path segment, so saved paths look like ~/.agents/output/<project>/markitdown/dqw4w9wgxcq/dQw4w9WgXcQ.md.
• Audio transcription in the verified 0.1.7 converter calls recognizegoogle; it is not local Whisper. Inspect the installed implementation and Microsoft's source before processing. A local/private-only request needs an actually local configured tool; do not send audio externally under a local-processing promise.
• Image OCR requires a configured capable backend. The Microsoft OCR plugin instructions require an llm_client and model; installing the plugin and passing -p alone is insufficient. Report unavailable OCR explicitly and use the documented integration only when its configuration and transport are authorized.
• Saved-output collisions fail before conversion if the destination already exists. Use another destination/input identity or obtain authorization to replace the existing output; slug naming alone does not preserve files.