Look ahead considered helpful

Typing is not the bottleneck

Here’s what AI actually speeds up: typing. Getting characters onto the screen. This was never the hard part.

The hard part was deciding what to type, and then making sure it was right — reviewing it, understanding it, answering for it when something breaks. AI doesn’t help with either. It just makes the easy part faster.

Every line of code that goes into production needs to be understood by someone. That was true when you wrote it yourself, and it’s true now that an AI wrote it for you. The difference is that you used to understand the code by default, because you’d just thought through the problem. Now you have to develop that understanding after the fact, by reading work that arrived fully formed.

At a five-person startup, review capacity is your scarcest resource. The founder who understands the whole system is also closing deals and talking to users. Every hour reviewing AI-generated code is an hour not spent on the things only they can do.

So the math isn’t “10x more features.” It’s: “We can generate 10x more code, but we can only review maybe 1.5x more.” Be selective about what you ask for.

Lookahead considered helpful

The problem with AI-generated code isn’t mainly that the AI doesn’t know what decisions you made three months ago. That’s a context problem — better prompts, longer memory, agent-written documentation. This will keep improving.

The deeper problem is that the AI doesn’t know where those decisions were leading. It doesn’t know the plan.

When a good senior engineer writes code, they’re not just solving today’s problem. They’re solving it in a way that makes tomorrow’s problem easier: “We don’t have multi-tenancy yet, but we will in six months, so I’ll structure this so it won’t need rewriting.”

But a good founding engineer is doing something different. They’re solving the problem in a way that makes tomorrow’s *growth* easier. Not tomorrow’s code — tomorrow’s company. They’re thinking about what happens when the feature succeeds and needs to scale, or fails and needs to die cleanly, or when the whole product pivots and half the codebase becomes irrelevant overnight.

AI doesn’t do either of these. It optimizes for the prompt. Ask for a settings page and it builds a settings page — a good one, probably. But it won’t ask whether you already have a pattern for settings pages, or whether this approach will make sense when you white-label the product next quarter.

One prompt, one solution, no lookahead.

I know the vibe-coding position: you don’t need to understand every line. Ship it, see if it works, move on. There’s something to this — the market will tell you soon enough if something is broken.

But *someone* still needs to understand it. Maybe not you, but someone you’ll hire in six months who has to debug it at 2 AM. Until we’re generating entire apps greenfield and throwing them away just as easily, the code you ship is code you live with. It accumulates. It has weight.

And the weight isn’t evenly distributed. Some features will live, most will die. Your entire audience will probably change. The startup comes into contact with the market and gets pruned — that’s the process. The question is whether the code that survives is code you can build on, or code that locks you into decisions you didn’t mean to make.

This is why lookahead at an early stage isn’t about writing durable code. It’s about not closing doors. Not letting bad infrastructure decisions, or accumulated slop, box you in when the market finally tells you what it wants.

Keep your options open. Make features easy to kill. Don’t let the speed of generation trick you into commitments you didn’t choose.

When you review AI-generated code, half the job is checking if it works. The other half is checking if it keeps doors open — can you delete this cleanly when it fails, extend it easily if it succeeds, or does it add another way of doing things you’ll have to reconcile later?

This is the hidden intelligence in fast-growing companies. Not shipping quickly — everyone ships quickly now. Shipping quickly *and staying light*. Moving toward something without getting weighed down by the code you wrote to get there.

AI gives you speed. Lookahead keeps you free. You need both.