Before you think about a problem, you prompt. That's not a workflow. That's a reflex replacing your first instinct.
Think about the last time you encountered a bug. Not a trivial one. A real one — something that required actually understanding what was happening.
What was your first move?
If you have been working with AI coding tools for 12 months or more, the honest answer is probably: you opened a chat and described the error.
That is not a workflow choice. That is a reflex. And it is worth examining carefully — because reflexes, unlike workflows, are not chosen. They are trained.
You used to have a different first instinct. You would read the error. You would open the relevant file. You would trace the stack. You might talk to a colleague. You would think about it in the shower, or on a walk, or while making coffee.
That instinct is not gone. It is just buried under a faster, more practiced response: prompt.
When you learn any new skill — driving, touch-typing, a musical instrument — there is a period where it requires conscious effort. Then the neural pathway becomes automatic. The action becomes faster than conscious thought.
AI prompting has followed exactly this curve for most engineers. The conscious decision to open a chat has become a pre-conscious reflex in less than 18 months of regular use. That is unusually fast even for highly practiced skills.
Productivity is not the issue. Prompts produce useful output. That is not in dispute.
The issue is what the reflex is replacing.
Problem appears → you think about it → you build a mental model → AI assists → your model is refined → understanding deepens
Problem appears → you prompt → AI produces output → you evaluate output → mental model stays where it was → understanding plateaus
The difference is not visible in any single interaction. Output gets produced. Features get shipped. Velocity looks fine.
The difference is visible in what is happening to your thinking. The mental model of your codebase — the felt sense of how things fit together, where to look, what to expect — is built in the 20 minutes of deliberate attention before you reach for AI. The reflex skips that 20 minutes. It also skips the understanding that was built there.
Over time: the model stops growing. The reflex gets stronger. The gap between what you know and what the AI knows widens, and you stop noticing it.
Stage 1: Deliberate choice. You decide to use AI. This takes some effort. You are aware of using a tool.
Stage 2: Automatic shortcut. AI is now your first stop for most problems. You still feel like you are making a choice. You are mostly unaware of what the alternative would look like.
Stage 3: Reflex, not choice. You no longer have a first instinct about small problems. The reflex fires before conscious thought. You have to actively stop yourself to do it the other way — and that feels harder, not because it actually is, but because the reflex is now the path of least resistance.
Most engineers who started using AI tools seriously in 2023 are now in stage 3 for most common problems. They do not notice this happening because the reflex feels like thinking.
For the next bug or problem you encounter — a real one, not a trivial one — try this:
Give yourself 20 minutes before you open any AI tool. Not to solve it without AI. Just to think about it first. Open the relevant files. Read the error. Trace the logic. Ask yourself what you expect to find and why. Write down your hypothesis.
Then open AI and compare: what you expected versus what it produces.
If the comparison surprises you more often than not, you are in the reflex stage. The AI is telling you things your mental model would have told you — if you still had the retrieval path to it.This is not about proving you do not need AI. It is about noticing whether your mental model is still being used — or whether it has been quietly replaced by the reflex to prompt.
The reflex is rebuildable. It just takes the opposite pattern of how it was built: deliberate, consistent replacement of the shortcut with the slower path, enough times that the slower path becomes familiar again.
The reflex is not inherently bad. Fast access to useful information is genuinely valuable. The problem is not that AI is fast. The problem is that when the fast path replaces the slow one entirely, something gets lost that the fast path cannot recreate: the understanding that comes from genuine struggle with a problem before you know the answer.
Engineers who maintain genuine expertise over time will be the ones who preserve both paths — the reflex for velocity, and the slower thinking for depth. Not because the slower path is always better. Because the slower path is the only one that builds the kind of understanding that makes the reflex smarter, not just faster.
The reflex is useful. Train it. But do not let it be the only tool you have.
Go deeper on how AI tools reorganize your thinking habits — and what to do about it: