Welcome to Verstreuen—meaning “to scatter”—where I unpack the ideas I’ve collected this week in my 🗃️ Zettelkasten, “note box,” personal knowledge management system. Here, I’ll share the highlights, insights, and stories I find interesting—and think you will too!

_{Forwarded this email? Subscribe here for more}

🗃️ This Week’s Highlights

This week's notes come from 135 new additions to the Zettelkasten—here’s the three that stood out most to share with you:

🟨 Why prompting AI is really just about good written communication

🟦 A simple framework to know when it’s time to quit your job (or stay and grow)

🟥 How graph-based agents aren’t just smarter systems—they’re mini organizations

I recently picked up Principles of Building Agents by Sam Bhagwat—a short but comprehensive read that distills insights from the front lines of startups working on AI. It’s a synthesis of practical patterns, technical frameworks, and learnings from the people actually building AI agents day to day.

Among the many takeaways, one core principle continues to echo through every chapter: all good prompt engineering is just good written communication.

That’s it. That’s the whole game.

Prompting isn’t some exotic new skill—it’s the modern evolution of how we write briefs, define roles, delegate tasks, and communicate expectations. The better your ability to express ideas clearly in writing, the better your ability to get results from AI.

This really clicked for me when thinking about it through the lens of an interface. Text isn’t just the interface between you and the model—it’s the interface in any human collaboration. We’ve always relied on written (or spoken) words to steer others. The difference now is that the recipient isn’t a person—it’s a machine.

The mental model I’ve been using since the GPT-2 API playground days is simple: “AI is just an intern you can only talk to through Teams chat.”

The intern might be brilliant, but if you give vague directions, lack examples, or fail to provide necessary context, they’ll just make assumptions and hand you something that’s “technically right” but completely misaligned.

It’s not that the model failed—it’s that you weren’t clear.

Getting the best output from AI isn’t about fancy architectures or clever hacks—it’s about creating clear, structured communication and treating your text like a design layer. Because whether you're prompting a model or briefing a teammate, the quality of the output reflects the clarity of the input.

**🗃️**

Lately I’ve been coming back to this really simple framework for thinking about work. It’s called the Triangle of Growth—and it’s one of those ideas that sounds obvious until you actually sit with it.

Not every job has to be a dream job. But if you're missing two of these three ingredients, it might be time to realign or move on.

But here’s the thing—I don’t think this is just about job satisfaction. I’ve been thinking about it more as a way to evaluate whether your current role is actually helping you grow, or just keeping you busy.

A job doesn’t need to check every box. But if it’s not helping you build something—skills, relationships, momentum—then you’re not compounding. And if you’re not compounding, you’re probably just treading water.

Every job is a platform. The real question is: Are you using it to build leverage?

Inspired by a video from Productive Peter, I’ve started thinking about roles in terms of how they help you:

1️⃣ Grow valuable relationships
2️⃣ Build stackable skills
3️⃣ Create visible impact

That’s what real growth looks like—where effort compounds into more opportunity. Putting yourself in the kind of environment where your effort multiplies. Where good work creates more good work. Where skills stack, networks expand, and reputation starts to carry its own weight.

• The people you work with – Are they challenging, inspiring, and helping you grow?

• The work you’re doing – Is it building skills that stack over time and create visible impact?

• The compensation you’re receiving – Does it support your ability to take risks, learn, and invest in yourself?

Not every job is going to have all three but if you don’t even have two, that’s a signal. Not always to quit right away—but to start thinking ahead.

**🗃️**

If you’ve played with ChatGPT or other large language models, you’ve likely hit the limits of single-agent setups. You send a prompt, get a response, and hope it works—but as tasks grow more complex, that approach becomes unreliable. One-shot prompting only takes you so far.

That’s where graph-based workflows come in.

Instead of relying on one agent to do everything, graph-based systems break tasks into smaller, specialized nodes—each focused on a specific part of the process.

Building agent graphs is like staffing a nimble company—you don’t need one rockstar, you need the right roles. Each node handles its piece of the puzzle and passes the result to the next. This creates a modular, multi-step system where information can branch, loop, or reroute based on logic or confidence thresholds.

In practice, this leads to more reliable, scalable, and reusable AI systems. You’re no longer betting on a single perfect prompt—you’re building a network of specialized agents that work together like a team.

And the more I use and think about designing these systems, the more it feels like we’re not just building workflows—we’re designing mini organizations.

Graph-based agent design now looks less like a flowchart and more like an org chart. Each node acts like a specialized teammate, and the overall system functions like a small company—with roles, communication paths, and escalation protocols.

Each node is a subsystem, specialized in some way, and aware of its teammates’ strengths. When a task falls outside its domain, it knows when to defer—routing information to other agents better equipped to handle that part. What you end up with is a collaborative, adaptive system of agents that can ask for help, pass the baton, and collectively move work forward.

The whole graph becomes a networked control system. You’re not just prompting—you're coordinating cognitive processes across multiple distributed components.

And that shift demands a new kind of design mindset. Less prompt engineering, more systems architecture. This isn’t about chaining prompts—it’s about crafting a system where prompts and agents work together coherently, dynamically, and intelligently.

What started as a workaround for brittle outputs is becoming a new design pattern—one that mirrors how we build real teams, real organizations, and real systems.

**🗃️**

Closing Thoughts

The more I explore this space, the more I realize that building with AI isn't just about technology—it’s about design at every level: systems design, communication design, and even self-design.

In a way, each topic this week—clarifying prompts, evaluating growth, architecting systems—comes back to the same idea: clarity is leverage. Whether you're shaping your words, your work, or your workflows, the ability to express intent with precision is the throughline.

The better we understand what we want, the better we can design for it—across tools, teams, or even time.

Thanks for reading Verstreuen

Thanks for taking the time to explore and reflect on my notes with me. If any ideas particularly resonated or challenged you, I’d love to hear your thoughts.

👋 Until next week.

-GH

What did you think of this edition? 📩
We’d love to know how this week’s Verstreuen landed for you!

Click to rate:
⭐ 1 - Not for me
⭐⭐ 2 - Could be better
⭐⭐⭐ 3 - Pretty good
⭐⭐⭐⭐ 4 - Really enjoyed it
⭐⭐⭐⭐⭐ 5 - Loved it!

Have feedback? Let me know—I'd love to hear from you!