Introduction
logbasset, my open source CLI for querying Scalyr logs, was already built to be friendly to AI coding agents. It ships a context command that prints an agent-facing reference — every command, flag, output format, time syntax, and even a list of flags that look plausible but don’t exist — plus a schema command that emits the same information as machine-readable JSON.
So when I sat down to add an installable agent skill, I had a fair question to answer first: if the binary already documents itself, what’s the skill even for?
Self-documentation isn’t discovery
The context command answers how to use logbasset. But it only helps an agent that has already decided to reach for logbasset in the first place. Nothing tells the agent “this tool exists, and it’s the right one when the user mentions Scalyr logs.” That’s a separate problem — discovery — and a context command can’t solve it, because the agent has to know to run it.
An agent skill fills exactly that gap. A skill is matched by its description, so it acts as the trigger layer: “when the user wants to investigate, search, or tail Scalyr/DataSet logs, use logbasset.” Once it fires, it hands off to the binary.
A thin skill, not a second copy of the docs
The obvious mistake would be to stuff the whole command reference into the skill. That instantly creates a second source of truth that goes stale the moment I add a flag — and now I’m maintaining the same documentation in two places.
So the skill does almost nothing. It tells the agent: check that logbasset is installed, then run logbasset context and logbasset schema to load the current reference, then build the query. It deliberately lists no flags of its own. A CLI change can never make it lie, because it never claims anything specific — it just points at the binary, which is always right.
That combination is the whole idea: the context command makes logbasset usable by an agent, and the skill makes it discoverable — with near-zero maintenance, because the skill delegates instead of duplicating.
Try it
If you use logbasset with a coding agent, install the skill alongside the binary:
|
|
It works with any skills.sh-supported agent — Claude Code, Codex, Cursor, and others — and the binary stays a separate install.
Conclusion
Giving your CLI a context or schema command is great: it lets an agent use the tool correctly. But a small, thin skill on top is what gets the agent to reach for it in the first place. If your CLI already self-documents, you’re most of the way there — a discovery layer is a cheap, high-leverage addition, as long as you keep it thin and let the binary remain the source of truth.