Skills, Context, and Agents. Oh My!

Happy Friday,

AI keeps changing quickly, and I wanted to try to break down a few concepts that folks talk around without really explaining them and why they make a difference. There is an interesting interplay between skills, context, and agents that is worth understanding and demystifying.

This will be based on Claude.

Context

This word gets thrown around with two meanings. The first refers to the set of information your LLM has access to, and the second to what it has in its memory. I want to talk a bit about this second definition.

All LLMs make use of context or "memory." When it runs out of memory, it will attempt to preserve what it believes to be important and purge the rest. Uniquely, it will attempt to preserve what started the session and where you currently are and then purge large chunks of the middle. When this happens, and I'm sure you've seen it, it is almost like it has a stroke, and you'll find yourself re-explaining things that it seemed to know just moments ago.

Learning to work with AI in ways that its context limits do not impact you is one of the most important skill sets to develop for longer, more complex working sessions. I still find that even daily users of AI have a weak understanding of this.

How do you think the software industry will do when it's using LLMs daily, while many are forgetting what's important and continue modifying code without any understanding?

Agents

Ok, now agents have changed their shapes over the year, but in essence they are a parallel instance of an LLM that operates with a unique set of instructions and unique context.

You can have an agent that acts like a developer, an author, an editor, or anything else you can imagine. You can build interesting operating models of inter-agent interactions by telling your agents how they must work with other agents and when.

For this newsletter, though, the most important thing is that each agent gets its own context. Using agents extends the shelf life of context because each agent has its own context. The more specialized the agent, the longer it will likely run before its context breaks. You'll likely use more tokens, but your work will be safer.

Skills

In the early days of Claude Code, there was just a Claude.md file that had all the instructions that every agent and LLM instance would follow at all times. The contents of that file would be put into context at all times.

If you wanted Claude to refer to you as "Your Highness," or, far more practically, to keep its answers terse and use one sentence instead of 3, putting these instructions in Claude.md would keep them alive at all times.

Unfortunately, for anything significant, your Claude.md file would keep growing with more and more information, and your available context would shrink because the contents of that file would fill it up by default.

So now we have skills. These are still pretty new. Skills are on-demand instructions that your agents or LLM can use and then forget. This lets LLMs get the benefit of what would have been a very comprehensive Claude.md but also allows it to be on-demand.

Imagine you're an author and you had a very large Claude.md file that had instructions on voice, grammar preferences, editorial passes, and more. If it were to help you write, it would struggle and hallucinate as its context collapsed under what was already written and attempted to write and edit more.

If you took those same elements and turned them into skills, you would be able to continue longer. You could turn each editorial type into a skill as well as an authoring skill and more. Then, as Claude works, it could figure out which skill to use at any given moment and leave out all the other unnecessary parts.

So What…

Getting more out of your LLMs seems to be an important skill today, and a big part of that is understanding how to manage the context of your LLMs and agents.

If you are interested in trying any of this out, start by working with Claude to develop the skills you need. Pay attention to how Claude will know to use them. From there, move into developing agents to stretch your context even further. Finally, develop patterns of documentation and reading documentation so Claude can more easily recover when its context pops.

I've been running one continuous session for three months across multiple devices without context rot failure.

That three-month run wasn't luck. It's a discipline, and it's the same discipline teams need to adopt AI without losing the plot. If that's a problem on your plate, reply and tell me where it's pinching. And if an engineering leader came to mind while reading, forward this their way. That's how my best conversations start.

Sincerely,

Ryan