Lately I've been really interested in watching how different people are structuring the way they work with AI. Not so much which tool they use, but how they organize the process around it. What they do first when a task is complex. How they break down the problem. How they verify what the model produces. How they incorporate mistakes to improve.
Last week I saw an interesting example shared by the director of Claude Code, where he explained his workflow. What caught my attention was that he wasn't talking about "magic" prompts or tricks to squeeze more out of the model. He was talking about something much closer to an actual work system.
Plan before doing anything
The idea starts before writing a single line of code or prompt. When the problem is complex, the first step is to plan: truly understand what you're trying to solve and what parts it's made of. It sounds obvious, but it isn't. The temptation to open the chat and start asking is enormous (and it's usually the first thing we do). And that initial planning is precisely what keeps the model from generating solutions without a clear framework.
Break it into small parts
Then comes something I found interesting: splitting the work into smaller pieces and assigning them to sub-agents. Instead of asking a single model to solve everything at once, you create small units that work on specific subtasks. Each one produces a piece of the final result, and that makes the process much more manageable and easier to review.
Verify everything
Once those pieces produce results, another phase kicks in that I think is key: verification. You don't assume that what the model generated is correct. You review it, cross-check it, re-run it if needed. It's a process that looks more like reviewing someone's work on your team than "using a tool."
Learn from what doesn't work
And finally, something I really liked: a continuous improvement system based on mistakes. Every failure or unexpected result becomes information to adjust the process. Over time, the workflow improves because it keeps learning from what didn't work before.
Reading it, I was left with a feeling I've been mulling over for a while: working well with AI looks less and less like "asking a model questions" and more and more like designing a work system around it.
A concrete example
Let's say I want to understand the market for restaurant reservation tools. Instead of opening ChatGPT and asking "what are the best tools?", the approach would look something like this.
First, planning. Before asking the model anything, I define what I need to understand: what companies exist, what their business model is, how much they charge, what problems they're trying to solve. That initial framework keeps the research from being chaotic.
Then, the breakdown. Instead of asking a single model for everything, split the work into more focused tasks. One agent can research competitors. Another can analyze pricing. Another can review user feedback. Each one produces a smaller piece of the overall analysis.
Then comes verification. The results the models generate aren't automatically taken as good. They're reviewed, compared against each other, cross-checked with external sources if needed.
And lastly, learning. If the model gets it wrong, if the information isn't useful, or if something important is missing, that feedback is used to adjust the process. Over time, you know what to ask for, how to split the work, and how to validate what comes back.
| Step | What it looks like |
|---|---|
| 1. Plan | Define the problem and its parts before opening the chat |
| 2. Break down | Split work into focused sub-tasks for separate agents |
| 3. Verify | Review, cross-check, and re-run — never assume the output is correct |
| 4. Learn | Feed mistakes back in so the next round is better than the last |
It's not about asking better questions. It's about designing better processes.
What I find most interesting about all of this is that it changes the way you think about working with AI. And that's where I think the most valuable conversation is: watching how different people are starting to structure their work around these tools, not just use them.