There's a right way and a wrong way to use AI.
The wrong, gross way: have Claude write and respond to all of your emails, write your blog posts, and communicate in your voice. (As Gen Z would say, this gives me ‘the ick.') It's the stuff AI slop dreams (nightmares) are made of.
The right way? Understand how the various components of a useful AI system work together. Then use them all intentionally.
Yesterday's episode of Intentional AI, the podcast I do with David Sparks, is about the various parts of an AI system. (You can listen to it on Relay and watch it on YouTube.)
Once you know them, you can use AI much more effectively.
We'll be going into each of these components in future episodes. But I wanted to plant a flag in the sand, so I have something to link back to going forward 🙂 Here are the five parts of an AI system that are the most useful to know.
1. The Harness
The harness is what you actually use and interact with—the software you can reach an AI model through. Example harnesses:
- The ChatGPT or Claude PC or Mac app (which connects to their cloud models);
- The ChatGPT or Claude phone app (which connects to their cloud models);
- A dictation app that polishes and then transcribes text for you (like Wispr Flow);
- An AI voice assistant, like Google Assistant or the upcoming Siri AI.
(As a side note: developers call all of the scaffolding around a model—from the app to the orchestration logic—the harness. I'm a fan of keeping things simple with the definition above.)
2. Context
Context is the next component—the information the AI keeps in its memory about the problem you're trying to solve. If you're using a chat-based AI, your context typically includes the entire conversation you're having with the AI—and also any text, images, or files you add to the conversation. If you're using Claude Cowork and you're working in a folder, you can pull files from the folder into the working context to get tasks done.
Context is such a big topic that I'll write a separate article going into how to think about and use it to advance your goals.
The thing to keep in mind for now: while it's getting better, AI currently has a fishlike short-term memory—it's very forgetful. What the AI keeps in mind at any one time is called its “context window.” This is akin to our brain's “working memory.” AI is constantly walking into a room, forgetting what it came in for.
3. Skill Files
The definition is in the name: a skill file is simply a set of instructions that tells AI what to do, and when.
For example, the skill files I use the most often are the ones that:
- Coach me toward my fitness and health goals by analyzing my workout data and food log photos;
- Coach me to become a better public speaker (by analyzing transcripts of my rehearsals);
- Help me manage my time;
- Update the e-ink home dashboard in my house;
- Are working with me and my wife to project manage a bedroom remodel.
To create a skill file, ask Claude (or ChatGPT or whatever you use) to create one for you. (I don't think I've ever written one from scratch—I've always written these alongside AI, as AI agents can structure them better than I can.)
Tell AI the problem you want to repeatedly solve. Then have it walk you through creating the file—as well as getting connected with the tools that will make it happen. Or have it interview you—and give it all the context you can think of to build the skill.
Tips for building a great skill file:
- Ask it to fortify the skill file after it makes a mistake or learns a more efficient way;
- Ask it often to reflect on its behavior: is there a better, more efficient way to do things?
- Get it to audit its skill file;
- Get it to do deep research as it builds the skill file.
4. MCP Connections
MCP stands for Model Context Protocol. The definition is in the name: it provides the Model that you're using with Context it can use via a standardized Protocol. Some examples of ones I have set up in Claude include:
- Chrome MCP: some AI agents, like Claude, can navigate Chrome for you—reading webpages, finding information, and filling out repetitive forms. I use this MCP connector all the time—for example, yesterday I had Claude navigate via Chrome MCP to find a new light for a bedroom remodel we're doing.
- Fastmail MCP: my email provider. This means my AI can access what emails I have in my inbox, alert me when new ones come in, and so on. I never have it send anything for me because I find that gross.
- Todoist MCP: the to-do application I'm currently using. I mostly use this to prioritize my day and analyze how my daily tasks fit with my broader goals.
- Home Assistant MCP: this lets me connect with my smart home setup directly from inside Claude.
5. Reference/Memory Files
To repeat from earlier, context is just any bit of data. So technically, memory and reference files are context too. But they should have their own section, because they have specific use cases.
Apps like Claude Cowork let you work inside a folder. This is when the app really comes to life, because you can keep running logs of certain things like:
- A task list you're working on together with the AI agent;
- A memory file of the actions you take;
- Long-term memory files of work you've done previously.
It helps to reference these files in the skill file—so that whenever the skill file is engaged, these things are being logged. As David mentions on the podcast: Skills + Reference = Memory.
Here's how the five fit together: context is the umbrella—everything the AI is holding in mind right now; skill files, memory, and reference files are the persistent pieces you load into that context; MCP is the standardized pipe that pulls in outside tools and data; and the harness is how you use it all.
Those are the five parts of an effective AI system! Here's where you can listen to this week's episode of Intentional AI, where David and I walk through all five.
