Don't Let It Do Your Thinking

01 / 04The Bite

This is about AI: the coding agents, the LLMs, the "agentic engineering" and "AI-native" transformation everyone keeps selling. I think about it like a vampire. I got that picture from a movie, Ryan Coogler's Sinners (2025), which I wrote about here. The vampire does not kill you. It bites you, and you turn into one too. Same face, same hunger. That is how the AI hype spreads.

I use Claude Code every day, and I build software for enterprises. I am writing this as an insider, and a critic.

The pitch is always pretty

You will be ten times faster. You will never write boilerplate again. The agent does it end to end.

That last line is the trick. The hard parts were the job.

02 / 04The Turn

The signs

You can spot the turned ones by what they post. Not software. Stories. The clever workflow. The framework that does everything. The weekend where an old system ported itself to a new language. Endless generated websites.

It looks great in a demo. It always looks great in a demo.

The mirror

Old stories say a vampire has no reflection. The turned engineer loses the same thing. Not the skill. The habit of looking at his own work and asking if it is good. If it is correct. If it is even his.

One trick is not a method

Then comes the worst sign. He takes the one thing that worked once, on one project, and calls it the way to build. He writes it up and tells everyone to do the same.

But the tool does not work the same way twice. If the result changes each run, you cannot treat one success as a repeatable method. You got lucky, and luck does not scale.

03 / 04The Dance

Boxing has a name for moving without an opponent: shadow boxing. The footwork looks sharp and the speed looks real, but nothing hits back. Demo projects are shadow boxing. Real projects are the actual dance. They hit back, and they hit back harder. New requirements, edge cases, migrations, the teammate who reads your code a year later.

04 / 04You Have to Invite It In

The oldest rule about vampires is simple. They cannot enter unless you invite them. AI works the same way. Nobody forced engineers to stop thinking, or to trust generated code they did not fully understand. They chose to.

That is the real line. Using AI to write code is not the problem. Using it to make your engineering decisions is. The tool is for execution: boilerplate, tests, the repetitive work. But the important parts still belong to you: the design, the tradeoffs, the review, and whether the solution is actually good. That judgement is the job.

Deterministic guardrails

You still need guardrails. AI is probabilistic. It gives a different answer every time. So safety cannot come from more AI. A rule in a claude.md file, a skill, a second model grading the first: AI checking AI is still guesswork.

The more deterministic steps you put in place before you reach for the model, the less you are leaving to chance. The probabilistic part should be the last mile, not the foundation.

Let the machine do. Never let it decide. That is when you get turned.