How to use the DataLad Data Management agent skill

What the datalad skill makes the agent do

Once installed, the agent knows to fetch real file contents before reading data, instead of treating a placeholder link as the file. It also knows to use datalad run and datalad rerun so each output has a machine-readable record of the command that produced it. Finally, it can decide whether DataLad is worth using over plain Git for a given repository.

The skill also covers publishing datasets to sibling repositories and storage remotes, such as a GitHub repository plus a storage remote.

• Fetch annexed file content with datalad get before analysis reads it
• Record commands and their inputs and outputs with datalad run
• Re-execute recorded commands with datalad rerun
• Publish a dataset to a Git sibling and a storage sibling
• Advise whether DataLad or plain Git suits a repository

How to install the datalad skill

Download the ZIP for the datalad skill from this page. Then choose one of two routes. For a personal setup, unzip it into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

The skill's own tool requirements still apply. You need Python 3.10+, DataLad 1.6.5, Git, and git-annex 10.x. Containers-run also needs datalad-container 1.2.6 and one of Singularity, Apptainer, or Docker.

• Personal: unzip into ~/.claude/skills/
• Project: unzip into the project's .claude/skills/
• Claude apps: Settings → Capabilities → Skills, then upload

Install DataLad and git-annex

The skill's reference install commands use uv. Install git-annex first if you can, from your system package manager or from PyPI, then install DataLad. Run the check command afterwards, because an old or missing git-annex causes many confusing errors.

The PyPI git-annex package supplies platform-specific binaries, and it does not supply Git itself. Configure your Git author name and email before creating or saving a dataset.

How to trigger the skill with natural requests

You do not need special syntax. Describe what you want in plain language, and the agent loads the skill when your request matches its trigger situations. These include cloning or fetching data from OpenNeuro, DANDI, datasets.datalad.org, or any DataLad dataset; a file that reads as a broken symlink or a small pointer; an analysis that needs a re-runnable record; publishing to a GitHub repository plus a storage remote; and deciding between DataLad and plain Git.

• Clone this OpenNeuro dataset and load one T1 image for me
• My NIfTI file says 'not a gzip file', check whether it is just an unfetched pointer
• Wrap this brain extraction step so the output can be rerun later
• Should this repository use DataLad or plain Git?
• Publish this dataset to GitHub with the data on a storage remote

Example: fetch data before reading it

A cloned dataset lists its files, but annexed content is usually not downloaded yet. Annexed files appear as symlinks into .git/annex/objects/, so a reader can fail on what looks like a real file. The skill directs the agent to run datalad get on the path first.

The example below is illustrative. Inspect the selected revision for the exact path, and install NiBabel before the Python read.

Example: record a reproducible run

datalad run executes a command and saves it in the same commit as its outputs, together with a machine-readable record. The example below is illustrative and requires bet and an existing derivatives/ directory. Declaring --input fetches the content before the command runs, and declaring --output unlocks the target so a second run does not fail on a read-only annexed file.

If the dataset has unsaved changes, datalad run refuses to start. Save or discard them first, or use --explicit to save only declared outputs. Check a command with --dry-run basic or --dry-run command before committing to it.

• Do not hand-edit the JSON run record in the commit message, because rerun parses it
• Save all input changes before claiming a run is reproducible

Safety notes: scripts and network access

This skill runs scripts and uses the network. Fetching and publishing move real data over the network, and remote data may require provider credentials. Local filesystem workflows work offline. Before fetching or publishing sensitive or protected data, check the access rules of that dataset.

The skill also lists Read, Write, Edit, and Bash among its allowed tools, so review what the agent runs, especially commands that save, drop, or publish data.

• Runs scripts: yes
• Uses network: yes
• Check access rules before handling sensitive or protected data
• Review publish and drop commands before the agent runs them

Tips for getting good results

Ask the agent to run datalad status --annex availability before diagnosing a file as corrupt. Many apparent corruption errors are unfetched pointers. Use datalad get on a directory rather than fetching files one by one.

Remember that git-annex whereis reads recorded state and does not contact remotes, so it shows what was last known rather than what is true now. Keep code-only repositories on plain Git.

• Confirm pointers with datalad status --annex availability first
• Fetch directories with datalad get sub-01/
• Use datalad wtf --section dependencies when behaviour looks impossible