It's a skill issue
Skills are an interesting emerging AI primitive. Essentially, a Skill is a a document which outlines instructions to an LLM or agent. Structured procedural knowledge. An instruction manual, to put it plainly. Simple, yet weirdly effective.
Skills are interesting precisely because they are beginning to look like a pretty viable middleware that can mediate between agents, humans and software. More sophisticated than a prompt, but just as simple as a “how to guide”.
Let me use a recent example of something I’ve built to illustrate this idea.
I wanted to build a GitHub-style “Contribution Graph” for my Substack posts, to get a visual sense of how often I was posting and where posts cluster during the week. So, I built a little tool for myself which did exactly what I wanted. Spat out a beautiful chart and some stats.
Then I thought, why not release this for other people to use too? The “tool”, is really just one html template which reads from a local csv file on disk, and a SKILL.md document which instructs an agent how to help a user get their real Substack data into that template. The repo (see below), is just those two files (and a README).
The interesting (and novel) thing here is how a user uses this tool. They don’t have to read the code, understand what it’s doing and follow instructions to set it up. They simply ask their agent of choice to look at the skill file and take care of the rest.
I’m noticing more and more software products using (and launching) skills. From official skills from the likes of Basecamp to tons of community-driven skills for Linear and Figma. As more and more products explore what a headless version of their software looks like (so agents can use it), Skills seems like a pretty reliable and viable building block in the emerging agentic stack.
PS. Check out the Substack Contribution Graph repository (and skill) here.