Issue #002: The Art of Thinking with AI

October 14, 2025

Welcome Back

Last issue was about why human in the loop matters. This one is about how to actually do it. Specifically, how to use AI as a thinking partner, not just a code generator.

Let’s dive in.

The Shift

LLMs are surprisingly good at extracting information from you if you let them. They synthesize, research, and summarize faster than you can. But most people treat AI like a vending machine: input prompt, get output, done.

There’s a better way. Work WITH the AI. Use it for planning, writing specs in your codebase, and thinking through business strategy. Let it be a thinking partner.

Here’s what that looks like in practice.

How It Actually Works

I needed to spec out a complex feature for Tether: the phone number purchase flow. This includes pricing display, error handling, payment integration, the whole thing. Instead of grinding through writing it myself, I tried something different.

I told Claude Code what we were going to build in a single sentence and then asked it to interview me. One question at a time, using each answer to inform the next. A dozen questions later, covering user flow, technical choices, success metrics, basically everything. The output? 424 lines of crystallized requirements. Ready for implementation.

I didn’t write a spec. I had a conversation that extracted everything from my head.

Or take this: recently I had a particularly productive day. I shipped five dozen commits across four projects in a single day. How? I spent the day creating specs and asking Claude to implement them. I worked on each project one at a time, no frantic context switching. The AI was multitasking for me. I stayed in the architect role: spec, review, next project. Each spec led to multiple commits as the AI handled implementation.

That’s the pattern now. Days where I ship a week’s worth of code happen fairly often. Of course there are still rabbit-hole days where not enough gets done; that’s software engineering. But the ceiling is higher than it’s ever been for me.

The Secret: Context Management

The trick isn’t just using AI more. It’s managing context; not too little, not too much.

I start with well-crafted Claude.md files in each project (and a global one in my user home folder). These “set up the world,” so to speak, with my preferences, tech stack, conventions, and project structure.

Then for each session, I mention relevant files or ask the AI to find them (with a little direction from me). For example, I might ask it to “look in the routes file for existing URL patterns when determining the URL for this new endpoint.”

Three mistakes I see others make:

1. Expecting AI to read their mind (too little context)

2. One-shot prompting without setup (lazy approach)

3. Overstuffing with irrelevant context (overwhelming the model)

Find the sweet spot. Give enough context for the AI to understand the problem space, but don’t dump your entire codebase. Claude Code and other specialized coding tools are also pretty good at finding what they need if you tell them to.

The Practical Workflow

Next time you have something complex to plan, try this:

Give the AI context. Use a voice tool like Wispr Flow and ramble. Tell it what you want to achieve, what’s on your mind, what you’ve done, what you want to avoid. Ask the AI to make a plan WITH you. One question at a time, each informed by your previous answer. Save the Q&A to a markdown file.

The results will be so much better than one-shotting it.

Yes, this takes longer per prompt than vibe-coding. But I think it takes less time overall. The results are dramatically better, and you get less back-and-forth during implementation. Plan once, let the AI cook with confidence. One or two iterations to perfection versus endless tweaking.

For small changes? Just prompt them out. Tiny bug fixes? Maybe faster to do it yourself. Match the tool to the task.

What Changed

I’m no longer a coder, really. I’m an architect, a product manager, and a code reviewer. I rarely write functions from scratch anymore.

I think about business goals, features, solutions, performance, UX, security. In other words, the big picture. But I still review everything because the code quality still matters, I’m just not writing it line by line anymore. This is the role transformation AI enables for solo developers. Strategic thinking over tactical typing.

Cool Things From the Community

Agent OS 2.0

Brian from Agent OS just shipped version 2.0, and it’s a solid improvement. I particularly love how it architects sub-agents for big features without compromising your main chat context. Check it out at buildermethods.com/agent-os.

Stripe Cheeky Pint Interviews

If you haven’t discovered this YouTube series yet, you’re missing out. Absolutely great interviews with CEOs of leading tech companies including most of the AI leaders. Real conversations that go deeper then I’ve seen anywhere else. Watch the playlist.

Build Crew

A community of developers shipping with agents. Great place to connect with others doing this work. Join at buildcrew.team.

Want to share your project here? Just hit reply and tell me what you’re building. I love featuring interesting work from the community.

Final Thoughts

That’s issue #002. The mindset shift from code generator to thinking partner changes everything. Try the Q&A approach next time you’re planning something complex. You’ll be surprised how much better the output is.

Keep shipping,
Joey

P.S. Hit reply and email me if you’re working with AI. I want to hear about your workflows and processes. Seriously, please tell me what’s working for you? What isn’t? Let’s talk.

P.P.S. If you found this valuable, forward it to someone who’s figuring out their AI workflow. They can sign up at jkudish.com/newsletter.