Skills

The Dagu skill is a packaged authoring reference for AI coding tools. It ships in the repository and covers which DAG type to pick, the step fields, the built-in actions, the harnesses, and the CLI. Reference files load per task rather than all at once.

llms.txt is the same material flattened into one file, for tools that read a URL instead of installing a skill.

Both target tools working on files. When a server is running, MCP is the better path: it supplies the reference itself and checks every edit before it is saved.

Install the skill

Install it with the GitHub CLI:

gh skill install dagucloud/dagu dagu

It carries the authoring rules that are easy to get wrong from general knowledge: when to reach for type: graph over type: agent, why id belongs on every step, when action: template.render beats a shell heredoc, and which file.* action replaces shelling out to cp or mkdir. It loads task-specific references on demand rather than all at once, covering step types, the CLI, Dagu Actions, harnesses, and build DAGs.

Run gh skill install --help for tool-specific installation targets.

Point a tool at llms.txt

The flattened file lives at:

https://raw.githubusercontent.com/dagucloud/dagu/main/llms.txt

It is generated from the skill sources, so the two never drift. Paste the URL into a tool that fetches context, or keep a copy next to the workflows in a repository so the reference is available offline.

Let the tool check its own work

Neither a skill nor a reference file guarantees correct YAML. Two commands close the loop, and both are worth putting in front of an agent explicitly:

dagu validate workflow.yaml
dagu schema dag steps

validate builds the DAG and reports the same errors the server would, without running anything. schema prints the accepted fields for any dot-separated path (dagu schema dag steps.container), so an agent can look up a shape instead of guessing it. Telling a coding tool to run dagu validate after every edit turns a plausible-looking file into a verified one, and the same command is what checks every YAML example in this documentation.

When to use MCP instead

A skill stops at the file. MCP covers the same authoring job and more, which is why it is the recommended path whenever a server is running.

	Skill or llms.txt	MCP
Configure	gh skill install dagucloud/dagu dagu	http://localhost:8080/mcp
Needs a running server	No	Yes
Authoring reference	Installed on the client	Supplied by the server
Validation	You tell the tool to run dagu validate	The server checks every edit before it is saved
After the edit	Nothing further	Starts the run, reads logs, debugs the failure

Use the skill when the tool edits files with no server to reach, or alongside MCP when you want the authoring rules present in the editor too. See MCP Clients for per-client setup.

Regenerating the reference

Contributors editing the skill sources under skills/dagu regenerate the flattened file with:

make llms

Both surfaces are checked in, so a change to the skill needs the regenerated llms.txt in the same commit. See Contributing.