Takeaway: The most powerful way to use AI is as a collaborative mind you design intentional systems around starting with a clear sense of what you're optimizing for. Estimated Reading Time: 5 minutes, 39s.
I recently gave a TEDx talk (I'll send you the video when it's out!), and I used AI to help me prepare for it. Not to write it—I designed the talk myself—but to do the grunt work that made the creative part possible.
I had Claude Cowork completely scan my newest book, Intentional, to pull out every single story and tactic from it. It ranked them by how interesting they were, and I used that as raw material to design the keynote. To rehearse, I then recorded voice memos on my iPhone, used an app called MacWhisper to turn the audio into timestamped text, and fed it back to Claude so it could coach me on pacing and emotional impact.
The day before the talk, I rehearsed twice. The first run was 13 minutes; the second was 16. I wanted to know why. Within seconds, it crunched the data and told me that I'd taken my time throughout the entire 16-minute version—it wasn't one section that went over, it was all of it. The pacing was just better. It's an incredible rehearsing coach.
I could focus on creating the art and creating an impact—not doing the drudgery. That's the promise of these tools when you use them well.
I'm personally finding that I use Claude Cowork a ton. Here are ten things I've learned about how to optimally use it! I've ordered these from least to most complex—depending on how often you already use the app.
1. Start with an intention.
As I write about in my newest book, intention has never been as important as it is today. Before you ask AI to do anything, know what you're optimizing for. With a new workout regimen, I started with the end goal in mind—longevity and cognitive performance, with fitness benefits as a side effect. With the TEDx talk, I optimized for how entertaining and helpful the speech could be. When you're working with agentic AI, you're manipulating and shaping intelligence. You need a north star for why you're doing what you're doing. And here's what's wild: we're not just guiding our own intentions anymore. We're guiding the intentions of agentic systems. Understanding intention and how it works has never been more critical.
2. Dictate your prompts.
I use an app called Wispr Flow to dictate my prompts instead of typing them. Over the last 54 days, I've dictated 132,000 words with it—maybe a third to a half of that was to Claude. It significantly reduces the friction of communicating with these systems. You just speak to them as if you're instructing someone, rather than typing out commands. Even though the robot isn't talking back, being able to communicate by voice feels more natural and makes the whole experience faster.
3. Have it interview you.
I have Claude interview me for a ton of stuff. It's one of the most enjoyable ways to use the app. To dot his, I had it design a “skill file” for itself to act as a world-class journalist—taking on the top journalistic practices so it can extract information optimally out of my mind. I use interviews to create skill files and to fish out ideas I wouldn't have surfaced on my own for articles like this one. Here's the thing about guidance work: it works both ways. We guide our robots into performing actions for us, and our robots can also guide us so they can extract optimal insights for the best approach.
4. Plan alongside AI.
Don't just ask it to do things. Take a step back and think: what is my ultimate intention here? Then plan alongside it. When I had Claude label a year's worth of business expenses, I didn't just hand it the CSV. I asked, “What approach would you recommend for this, and is this in the optimal format for you to do a good job?” Plan together, provide feedback on the plan, and guide it toward your intention before anyone takes action.
5. Always mind your agent's context.
My friend David Sparks describes these robots as very intelligent employees with very bad amnesia. Part of managing that amnesia is understanding their context—their working memory. You need to know what files it's working with, what skill file it has loaded, whether there's a memory file tracking your goals. If you don't prepare it properly, you're going to get frustrated. Think of it like onboarding a brilliant new hire who knows nothing about your specific situation.
6. Collect context wherever you can.
The richer the context you give these systems, the better they can perform. For example, I use an app called Health Auto Export for the iPhone that exports my workouts and health metrics to a CSV file to Dropbox at regular intervals. That same Dropbox folder is the working folder for my health coaching conversation—so the AI always has fresh data to coach me on. Design your folders as workspaces for a mind.
7. When something goes wrong, fix the system—not the robot.
I used to get frustrated when my health coach wouldn't update my dashboard after a workout. But it wasn't the robot's fault—updating the dashboard simply wasn't part of the skill file. When something doesn't happen, it's not because the robot failed. It's because the system you designed didn't account for it. Adjust the system.
8. Calibrate supervision to risk.
The greater the risk, the greater the supervision. For example, when Claude was updating my WordPress website, I watched it do every single step. When it was labeling my expenses? I just scrolled through the results afterward and spot-checked. For low-risk tasks, double-check the results, not the process. For high-risk tasks, observe the process live.
9. Think of your AI as a mind, not a tool.
Your AI's “context” acts as its working memory—what I call in Hyperfocus your “attentional space.” Its skill files are its knowledge and habits. Its memory files are its long-term memory. Your prompts are its deliberate intentions. When you shift from thinking of AI as a calculator to thinking of it as a mind, everything changes. You stop “using” it and begin collaborating with it. You design working environments for cognition, not just input fields for commands.
10. Think of every task as an optimization problem.
Everything is an optimization equation with these systems. You're always trying to figure out the right variables to optimize for. When I had Claude build a journalism skill for conducting interviews, or a health coach skill customized to my goals, the process was the same: have the AI reflect on what the top variables are to optimize for, and then what the most optimal ways to achieve them are. Then it designs the skill itself—customized to what you actually want to accomplish—starting with your intention.
