How do you stay sharp when AI does most of the work for you?

How to keep your skills sharp while using AI every day

You stay sharp by paying a deliberate tax on the work AI does for you: redo one sub-problem manually each day, force the AI to defend its reasoning before you accept it, and spend ten minutes on spaced-repetition drills of fundamentals. Naval's idea of specific knowledge still applies — except now AI evaporates the surface layer faster than ever.

Skill rot from AI is real, it is measurable, and I have felt it in my own work this year. There is a quietly viral Show HN this week — Fata, a spaced-repetition app explicitly framed as "to fight skill rot from AI coding" — and it landed on Hacker News with 100+ points and a comment thread full of senior engineers admitting the same private worry: when an AI writes the SQL, the regex, the migration, you forget how to write the SQL, the regex, the migration. The same week, LWN ran a piece called "AI agent runs amok in Fedora and elsewhere" about agents committing changes their human reviewers no longer understood well enough to catch. Both stories point at the same loss, just at different altitudes.

I have been using Claude every working day for a year and a half. I am an AI architect, a coach, a founder — three jobs where being sharp is not optional — and I will tell you plainly that I caught myself last quarter unable to write a reasonably complex Postgres query from memory that I could have written cold in 2024. That is not a moral failing. That is the cost of constantly delegating the first draft. Naval Ravikant calls specific knowledge "knowledge that can't be trained for" — the stuff you actually own that lives in your hands and your gut. AI evaporates the surface layer of that knowledge faster than anything I have seen, because it is so good at being good enough. The honest question is not whether to stop using AI. The honest question is: what is the deliberate tax I pay so the floor underneath me does not erode?

What I do, in practice, is three things, and none of them are heroic. First, I redo one sub-problem by hand every day. If Claude wrote a SQL query for me at 10 a.m., I rewrite a different one — sometimes the same one — without help at 4 p.m. The rule is one rep, daily, on the thing I am most tempted to outsource. This is how I learned guitar; it is how surgeons keep technical skills under anesthesia coverage; it is closer to what athletes call "deliberate practice" than to learning. The reps are the price of admission to keep using the tool.

Second, I make the model defend itself before I accept its work. Ethan Mollick, in Co-Intelligence, calls this treating the model as a colleague — but specifically the kind of colleague you red-team. I paste back its own output and ask "what is the weakest argument here, and what would a senior engineer ask in code review?" Roughly a third of the time, the model finds a real problem in its own reasoning. The other two thirds, I am the one who finds the problem while reading the critique. Either way, my thinking is in the loop. This is the single highest-leverage habit I have adopted, and it costs me about three minutes per turn.

Third — and this is where Fata's instinct is exactly right — I run a small spaced-repetition deck on fundamentals I no longer touch every week. Mine is fifty cards: SQL window functions, three regex patterns I always forget, the difference between covariance and correlation, two specific Claude prompt patterns, a handful of coaching frameworks (Whitmore's GROW questions, the Co-Active four cornerstones). Ten minutes a day with the deck. The point is not to memorize trivia; it is to make sure the words still land in muscle memory when I need them at 2 a.m. with no internet.

The trap to name out loud is the smooth feeling. When AI is doing the work, your day feels productive and frictionless in a way that, paradoxically, should worry you. Cal Newport's Deep Work argument was that the discomfort of effort is the signal something valuable is being built; remove the discomfort entirely and you are building nothing. I treat the smoothness as a yellow light. If a week goes by where I have not been confused by anything I worked on, I have not been learning anything — and the moat that pays my mortgage has shrunk by another millimeter.

One last honest note: you cannot win this by going back. The people I know who quit AI entirely are slower, not sharper. Sharpness now means using the tool aggressively and paying the tax religiously. Fata users on the HN thread had the same conclusion. It is not a victory lap. It is a daily decision to stay in the work.