The most productive tension in product teams is not between engineering and design. It’s between maximizers and focusers.

Maximizers lean toward speed, optionality, parallel bets. Focusers toward depth, coherence, doing less for more.

Neither is right. Both create value and risk.

This explains why some conflicts feel unresolvable. A maximizer sees a focuser killing momentum. A focuser sees a maximizer burning trust for short-term gain. Different languages for what “good” means.

I am a maximizer. My instinct is “why not both?” Parallel bets surface opportunities you miss with pure focus. But maximising without constraint creates shallow progress everywhere and deep progress nowhere. The art is in knowing when to switch.

The smartest person in the room is not always the most valuable. The person who actually does something is.

High agency is the belief that you can act, combined with the will to follow through. Seeing a problem and deciding: I’ll solve this. Without waiting for instructions, without asking permission for every detail.

Intelligence without action is potential energy that never converts to movement. You can be brilliant and achieve nothing. You can be averagely intelligent and get an enormous amount done.

Why is agency so rare? We have spent decades optimising for compliance. Schools reward correct answers to set questions, not the asking of interesting questions. Companies bought obedience by the kilo. Follow the process, do what you’re told, don’t deviate.

That trained agency out.

The people who get ahead are the ones who train it back in. Who choose bias toward action over analysis. Who ship and then improve instead of planning endlessly.

There is a difference between prompting until something works and engineering with AI assistance.

The first is hacking. You try things until they work, copy code you don’t understand, and hope it keeps working. That’s fine for experiments and prototypes. It’s not how you build production software.

The second is what Willison calls “vibe engineering.” You use AI as a tool, but the engineering fundamentals remain. Writing tests. Reviewing code. Understanding what you ship. Taking responsibility for the result.

The irony: AI tools make engineering fundamentals more important, not less. If you can generate code faster, the bottleneck becomes validating that code. And validating requires understanding.

Engineers who know their craft get leverage from AI. They can do more, faster. Engineers who rely only on AI without the fundamentals build houses of cards.

AI coding tools promise 10x speed. What they often deliver is 10x more code to understand and maintain.

The trap is not in the AI. It’s in how we use it. If you treat AI as a code generator that produces output, you get exactly that: more output. More files, more functions, more lines. But more code is not the same as better software.

The problem shifts. Instead of writing code, you are now reviewing code. Code you didn’t write yourself, following patterns you didn’t choose, making assumptions you didn’t validate.

The paradox: the faster you generate code, the more time you spend understanding what was generated.

What works is using AI as a thinking partner, not a typist. Let it reason about architecture, edge cases, alternative approaches. The value is in the dialogue, not the output.

The difference between managing and leading shows up under pressure.

Managers retreat into control mode. More rules, more checkpoints, more reporting. It feels productive but signals distrust.

Leaders step into trust mode. They explain the situation, ask for input, and give people room to act. It feels vulnerable but builds ownership.

The sharpest contrast: information. Managers treat details as a tool for power. The less others know, the more dependent they are. Leaders do the opposite. They share everything relevant and trust people to handle it well.

What it really comes down to is discomfort. Difficult conversations, bad news, conflicts. Managers avoid them or wrap them in processes. Leaders lean into them, directly but with care. “This is hard, but I care about you, so let’s talk.”

The bottom line: don’t avoid the hard things.

ChatGPT remembers things about you. It builds a profile from your conversations and adjusts future responses accordingly. Convenient, but also a little unsettling when you don’t know exactly what it’s retained.

Claude does it differently. Every conversation starts fresh. No automatic profile, no invisible personalisation. If you want Claude to remember something, you need to ask explicitly.

The design choice is interesting. Most AI companies optimise for convenience. Anthropic seems to optimise for predictability and control.

In a professional context that is more valuable than it sounds. You want to know what the model knows. You don’t want to be surprised by assumptions based on a conversation from three weeks ago that you’ve forgotten.

It’s a trade-off. Less magical personalisation, more transparency. For casual use ChatGPT’s approach is more pleasant. For work where consistency and control matter, Claude’s approach feels better.