You don't build the airplane in the sky. You design it in the office before it ever reaches the runway. But the startup culture of “move fast and break things” loves to cut corners - and that is super F*ing expensive with AI - and many people, are finding out the “fun” way - including me. Let’s get into it - so we can get over it.

There is an old military saying:

Sweat more in training, bleed less in battle.

It’s a simple equation. The more you sweat in training, the less you bleed in battle. For the AI era, I'd rewrite it like this:

Sweat more in chat. Bleed less in rework.

Yes, that usually means you'll bleed less in tokens too. But tokens aren't the real problem.

Rework is.

Recently I found myself refining Heroik's AI Brain inside a swarm of agentic workflows when I hit my usage limits. It wasn't because the AI was doing anything wrong. It was because I had wandered into the most expensive phase of AI work while I was still discovering the problems.

That experience exposed something I think many people are about to learn. AI hasn't reduced the cost of bad thinking. It has amplified it.

The Dangerous Illusion AI Creates

One of AI's greatest strengths is also becoming one of its biggest traps.

Execution has become almost frictionless.

Need a proposal? Done.

Need software? Done.

Need a presentation? Done.

Need a website? Done.

Need a research report? Done.

Because it's so easy to begin, people begin before they're ready. They start building before they've finished thinking.

They sketch while they're coding.

They architect while they're deploying.

They discover requirements halfway through implementation.

It feels incredibly productive because something is always happening.

But motion isn't necessarily progress. Flapping your arms is not flight.

It's just expensive learning happening later than it should.

It turns out building airplanes while flying them is expensive and often just plain stupid.

AI Didn't Eliminate Design

It Elevated It. Historically, execution was expensive.

• Blueprints had to be finished before construction.
• Factories had to be designed before production.
• Software required architecture before coding.

Those constraints forced discipline.

AI removes many of those constraints. And ironically, that makes discipline even more important.

When execution becomes cheaper, the value of good thinking increases. Not because thinking changed. Because the cost of changing your mind after execution has begun is still enormous.

Every Unknown Becomes Another Iteration

Every project begins with uncertainty. That's normal. The question is when you pay for re-solving it.

Imagine you're building a pricing framework for a client.

You ask an AI agent to create one. It does.

Then you realize it doesn't account for distributors.

So you revise it.

Then you discover municipal utilities behave differently. Revise again.

Then sales wants pricing flexibility. Revise again.

Then procurement introduces supplier constraints. Revise again.

Then leadership wants margin guardrails. Revise again.

Nothing went wrong. The work simply happened during execution instead of before it. Every unanswered question became another iteration. Every iteration became another prompt.

Another context window. Another review cycle. Another round of compute.

Another hour of attention. Gone.

The work didn't disappear. It merely moved into the most expensive part of the process.

AI Compounds Design Debt

Software teams talk about technical debt. AI introduces something just as dangerous.

Design debt.

Design debt is every unresolved assumption you carry into execution.

• Every vague objective.
• Every missing edge case.
• Every undefined success criterion.
• Every requirement you "figure out later."

AI is incredibly good at executing. It is also incredibly willing to execute a mediocre plan.

The better the execution engine becomes, the more expensive poor design becomes. That's because AI compounds design debt at machine speed. It doesn't reduce the cost of bad thinking. It accelerates paying for it.

The Heroik Loop

This has become one of the operating principles we're adopting. Essentially it’s getting back to basics with an AI twist.

1. Sketch

Capture the idea. Don't worry about perfection. Just get it out of your head. Sketch it in your notes with Chat.

NOTE: by “Chat” I mean don’t put that AI in “work” mode. That instantly triggers the bias to start building and shipping things - and it goes through tokens faster than a turd through goose.

Beat up your sketches. Sketch over them. Iterate the sketch. “Chat” is much cheaper than “Cowork” or work or build mode, or handing it off to an agent biased to “execute”.

2. Challenge

Ask uncomfortable questions.

• What are we actually solving?
• What did we miss?
• What assumptions are hiding?
• What would someone who disagrees say?

Pro-Tip: You can almost one shot this with one of my favorite prompts. Make chat solve this by asking it “What did you miss?” a few times - greatly improves the output.

Check the problem with chat. Here’s a good question to explore with chat.

• Is this even the right problem?

3. Red Team

Pressure test the idea. Look for failure before reality finds it for you.

Ask:

• How does this fail?
• What's brittle?
• Where will people misunderstand?
• What happens six months from now?

Don't protect your idea. Attack it.

Have chat attack it. Embrace thinking mode. It’s cheaper and AI is a decent sparring partner.

If it survives criticism, it deserves execution.

4. Refine

Clarify the language. Reduce ambiguity. Remove unnecessary complexity. Strengthen the weak spots. Repeat until the idea stops changing dramatically.

After finding out how expensive design debt is (as it cascades into tech debt and sucks tokens faster than a digital girlfriend at a nerd convention),

5. Write the Specification

Freeze the important decisions. Execution should solve implementation problems. It shouldn't still be deciding strategy.

6. Deploy

Now let the agents work. This is where AI shines. Not because it's thinking for you. Because you've already done the thinking.

Chat is the Whiteboard

Many people think chat is the lightweight version of AI. I think it's the opposite.

Chat is where the expensive work happens.

It's where:

• ideas are challenged.
• assumptions are exposed.
• strategies are compared.
• second and third order effects become visible.
• ideas become specifications.

That's incredibly valuable because conversation is still one of the cheapest places to discover you're wrong.

Once agents begin executing, every new discovery becomes rework. And the tokens suck it up.

Vibe Coding Isn't the Problem

Vibe coding has its place. Rapid experimentation is powerful. The problem is when people begin vibe architecting.

They're designing the product while simultaneously building it. The AI happily keeps generating. The human happily keeps revising. Everyone feels productive. Meanwhile, design debt quietly compounds. And the AI token costs explode!

The issue isn't that AI moves quickly. The issue is that humans stop slowing down long enough to think, and disconnect from the token costs and downstream effects.

The Real Competitive Advantage

The winners in the AI era won't simply be the people who execute the fastest. Execution is becoming a commodity. The advantage will belong to those who reduce rework before execution ever begins.

The advantage favors the disciplined - those who sweat more on the mats, the whiteboards, the design process.

• The people who ask better questions.
• Challenge assumptions earlier.
• Design more intentionally.
• Write clearer specifications.

And only then unleash machines to execute.

Because AI is an extraordinary amplifier. It amplifies clarity. It also amplifies confusion. The choice belongs to us.

Heroik Principle

Sweat more in chat. Bleed less in rework.

Every unanswered question becomes another iteration.

Every hidden assumption becomes another revision.

Every vague objective becomes another context window.

The cheapest place to discover you're wrong is still the whiteboard.

So sketch longer. Challenge harder. Red team relentlessly. Write the specification.

Then let AI do what it does best. Build. Because the goal isn't just to execute sooner anymore. It's to execute once -  to get it mostly right the first time.

This sounds like standard practice and basic discipline for the pros, but AI brings a broader audience ot the process, who may not have those disciplines yet. And even for some of the pros, AI vibing is tempting to skip those disciplines. So may this serve as a reminder for the pros, a saving grace for the vibe addicted, and decent meat for newbies.

Hopefully it inspires you to at least factor for your design debt and technical debt.