Markitdown

markitdown

Convert a source file to Markdown, save it next to the source, and — only with explicit approval — clean up formatting artifacts left over from the conversion.

This skill wraps markitdown, Microsoft's open-source CLI for converting files to Markdown — all credit for the actual conversion goes to that project. This skill only adds the install-and-fix-up workflow around it.

Step 1: Resolve the input

The input may arrive as an attached file, a bare filename, or a full path.

• If it's a full path or attachment, use it directly.
• If it's a bare filename, check the current directory first, then common

locations like ~/Downloads and ~/Desktop. If you find more than one match or none, ask the user rather than guessing — picking the wrong file is worse than asking a quick question.

Step 2: Make sure markitdown is installed

Run scripts/ensure_markitdown.sh. It checks whether markitdown is on PATH, already installed at ~/.local/bin/markitdown, or importable as a Python module even without a console-script entrypoint (covers venvs, system package managers, and non-standard pip installs) — and only installs it if none of those find it. No need to ask before installing. This mirrors how it was set up the first time, so re-running it on a machine that already has markitdown is a fast no-op.

The install path is platform-aware:

• macOS uses Homebrew (python@3.12, pipx), the reliable standard there.
• Linux (including WSL, which reports as Linux) prefers pipx/pip3/pip

directly and never requires Homebrew, since most Linux users won't have it.

• Native Windows (Git Bash/MSYS/Cygwin, not WSL) looks for python3,

python, or the py launcher, uses pipx if present or pip install --user otherwise, and checks %APPDATA%\Python\PythonXY\Scripts for the installed executable — pip's Windows install location, which isn't ~/.local/bin.

If the script still fails after that, surface the error to the user rather than trying alternate install methods — the failure usually means something about the machine needs their attention first (e.g. no pip available at all).

Important: each shell command you run is typically a fresh process, so a PATH fix made by the script doesn't carry over to the next command. Source the script (source scripts/ensure_markitdown.sh) in the same shell invocation as the actual conversion in Step 3, rather than running it as a standalone step — that way the corrected PATH is still in effect when markitdown is invoked.

Step 3: Convert

The output path is always the source file's directory + the source's base filename + .md — i.e., convert it in place, next to the original. This keeps converted files discoverable and colocated with what they came from, the same way document.pdf and document.md sit side by side.

Before writing, check whether that .md path already exists. If it does, ask the user whether to overwrite it, save under a different name, or skip the conversion — don't guess, since an existing .md file next to the source might be someone's prior edits, not just a stale artifact.

Run the conversion:

markitdown "" -o ""

If the command fails, report the error plainly and stop — don't attempt the review step on a conversion that didn't succeed.

Step 4: Review pass (only after a successful conversion)

Markitdown's extraction is mechanical, so certain artifacts show up predictably, especially from PDFs and Office documents:

• Fragmented sentences — lines broken mid-sentence or mid-list-item

because the source PDF wrapped text at a fixed width. Rejoin these into proper paragraphs and list items.

• Missing headings — a document's title and section names often come

through as plain paragraphs. Look for text that's clearly acting as a title or section label (short, standalone, followed by body text) and promote it to the appropriate #/##/### level.

• Ligature and encoding glitches — things like fi instead of fi, or

mismatched smart quotes.

• Flattened tables — a table in the source sometimes comes out as loose

runs of text with no visual structure. Reconstruct it as a proper Markdown table or list if the row/column structure is still recoverable from context.

• Likely typos — obvious spelling mistakes that look like they came from

OCR or extraction noise rather than the author's intent.

Read the converted file yourself and identify which of these apply. Then show the user a clear before/after (a diff, or a summary of the specific changes) in the chat — do not touch the file yet. The point is to let them see exactly what's changing before it's permanent, since automated cleanup can occasionally "fix" something that was actually correct in the original.

Step 5: Apply only on approval

If the user approves, write the cleaned-up version to the .md file. If they want adjustments, iterate on the proposal first. If they decline or don't respond affirmatively, leave the raw conversion output untouched — the unedited markitdown output is still a valid, usable result on its own.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ErickRamirezAU
• Source: ErickRamirezAU/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.