A couple of years ago, every new AI demo put a knot in my stomach. I write software for a living. The models were getting faster at the part of the job I had spent years getting paid for. The question was ugly and simple: if they can write the code, what am I for?
The answer was not "type faster." The valuable part was never the mechanical act of producing lines. It was breaking the problem open, choosing an approach, and living with the consequences. The model can draft. It cannot own the decision.
That distinction is how I think about career value now. Three layers: what you Do, what you Know, and what you Think. AI can help with all three. It cannot replace the stack.
Do, Know, Think
Do: create, validate, automate
Doing is the visible work. With a model in the loop it should get sharper, not sloppier.
- Create. Blank-page work is the easy win. First drafts, boilerplate, the email you have been avoiding. Let it put 80% on the page. You still decide what survives.
- Validate. Reviews, proofreading, the second set of eyes you never have enough of. Use it to scan. Do not use it as a substitute for tests you understand.
- Automate. Not just the dumb loop. Workflows that notice when something looks wrong and react. The point is a system that does not need you to click every time.
If "doing" is the only layer you hand to the model, you have built a faster pair of hands and left the job the same size.
Know: research, summarize, connect
Knowledge is what makes the doing worth shipping.
- Research. Ask a narrower question than you would type into a search bar. A good prompt is a good research assistant.
- Summarize. Long reports, ugly stack traces, a folder of notes. Compress them so you can think, not so you can skip thinking.
- Connect. The useful move is the non-obvious one. Different viewpoints, adjacent domains, the dependency you did not see. Personas, a second pass from another angle, a "digital twin" of the system (Gems and tools like it) — anything that shows you the relationship you missed.
Think: question, envision, strategize
This is the layer that still pays.
- Question. Put your own material underneath the prompt — requirements, configs, the internal wiki. NotebookLM, Eesel, and tools like them are useful here because the context is yours. The quality of the answer follows the quality of the question.
- Envision. Generate options. Run the ugly version of the future. This is not a prophecy. It is a cheaper way to see the shape of a bet.
- Strategize. Sequence, risk, what to do first. A plan the model wrote is a draft. The call is still yours.
They compound
Doing is fuel. Knowing is whether the fuel is aimed at anything. Thinking is which mountain you pointed at.
Each layer multiplies the one below it. Better knowledge makes the same hours of work hit harder. Better thinking makes sure you are not becoming excellent at the wrong problem. Strap a model onto all three and you get speed. Strap it onto "do" alone and you get a lot of output you cannot defend.
I used to treat AI as a way to offload the boring parts. That is the smallest version of the tool. The larger version is a partner for the parts of the job that used to take a week of reading, a whiteboard, and a guess.
Copilot can put the function on the screen. It cannot tell you whether that function should exist. That is still the job.
What to actually change
Look at your week.
Where does "doing" eat time you could spend deciding? Where is your knowledge a pile of tabs instead of a picture? Where are you reacting because you have not asked a better question?
Use the model there. Let it take the toil. Keep the judgment.
The career that survives this is not the one that types the most. It is the one that can still say why the work is this work, and then use every available tool to do it well.