Skills are somewhat misleadingly named, in my opinion. But I guess “atomised, contextual prompts” is a bit less catchy.
We’ve seen lots of examples of skills that relate to software, coding styles, design, open source libraries, et cetera. But if you think of them as atomised prompts, the design space opens up pretty quickly.
Here are a few examples of types of skills I’ve been using in my day-to-day.
Informational / “glossary” skills
Skills are a way to propagate context to an agent at a contextually relevant time.
…That’s it!
A perfectly valid use of a skill, then, is simply to define a term. Agents can reach for that definition when they encounter the term and aren’t sure what it means.
For example, I have a global skill called tessl-definition. Its description is straightforward: “Definition of Tessl, available to agents who are unsure what it means.” Whenever I mention Tessl in a new session, the agent can look up exactly what Tessl is and what we do. This is especially useful because much of the relevant information about Tessl post-dates common training cut-off dates.
It also means I don’t have to re-explain the same context over and over again. If something changes, I can update the skill and increment its version.
You can probably imagine many ways this could be useful to an individual or a team. “Production-ready” might have a specific meaning at your company, for example. An agent might reasonably assume it means tests and lint are passing and the code builds. But what about user-facing documentation? README updates? The state of the branch? Or, yes, the obviously LLM-sourced emojis in your code?
Glossary-style skills can also be useful for high-level system overviews. If I mention the mercury repository, I’d like my agent to have a rough idea of what I’m referring to—without trying to grep my entire disk. In many cases, a short overview is enough. This is another place where informational skills work well.
Behavioural skills
Behavioural skills describe how an agent should behave in certain situations. They’re playbooks for reacting to users, and situations.
A clear use case for a behavioural skill is when an agent needs to respond to a frustrated user. I wrote a skill which provides agents a clear path forward when users are starting to lose it (not that I'd know what that's like). It triggers quite reliably:
Behavioural skills can cover a wide range of situations, such as:
- Handling frustration, confusion, or conflicting feedback
- Staying in a specific working mode, like brainstorming or teaching
- Reacting to specific types of output from tools
- Providing vibe-based guidance depending on the user’s tone or goals
These skills act like little playbooks. And importantly, they help prevent default behaviours that might be unhelpful, or even destructive.
Compositional / Routing skills
Did you know that a skill can route to another skill?
A pattern I’ve found useful is an overview skill that routes to more detailed ones. For example, a <topic>-101 glossary-style skill can provide a high-level explanation, then explicitly suggest other skills for deeper dives. This keeps individual skills small and focused, and avoids loading too much context at once.
Other routing-style skills help classify a situation before acting. A skill might help the agent decide whether it’s dealing with a bug or a design question, or whether the user is looking for code or an explanation of something. Making these distinctions explicit often leads to better outcomes.
Routing doesn’t have to happen immediately. An overview skill can also describe when it makes sense to load other skills later on. This can be useful if more advanced skills aren’t triggering when you expect, or if they require context that hasn’t been established yet. In some cases, you may only want a deeper skill to trigger if the overview skill turns out to be insufficient for the task at hand.
In the example below, I have a skill describing some basic criteria for personal website UIs. Nothing too prescriptive, just basic guidance. At the end of the SKILL.md (http://SKILL.md), I’ve included:
“At the end of your review, make sure to remove signs of vibe coding by using the de-llm-ui skill.”
This differentiation helps keep my context organised. I don’t have to remember to ask the agent specifically to check the site for signs of vibe coding - rather, I can encode that expectation into the entrypoint “review” skill.
Remember, skills are simply adding to your context window whenever they are loaded. So, skills with information about other skills can be highly effective.
Examples of this kind of skill include:
- Overview skills that point to more advanced topics
- Skills that help classify a problem before solving it
- Context management helpers that limit scope or verbosity
These skills are particularly useful for keeping agents focused and for managing limited context windows.
Closing thoughts
Once you stop thinking of skills as “things an agent can do” and start thinking of them as reusable, structured context, they become much easier to design intentionally.
Most of the skills above are just carefully written prompts. The difference is that they’re named, reusable, and available when they’re actually needed.
That small shift in perspective opens up a lot of room for creativity.
Have a play with Skills using Tessl (quickstart here) to manage and share them, we would love to try them out!
