Nobody sets out to forget how to code. It happens the way most skill loss happens: one convenient shortcut at a time. You used to write the regex yourself; now you describe what you want and paste in whatever the model gives you. You used to trace a stack trace line by line; now you drop it into a chat window and read the summary. Each substitution is individually rational. The sum of them, over a year of AI-assisted work, is a developer who can ship features fast with the tool on and freezes a little when the tool is off.

That's the uncomfortable subtext underneath the vibe-coding boom this month — Lovable reportedly closing in on a $13.2B valuation, Replit and Factory each raising into the billions, SpaceX undercutting frontier pricing with Grok 4.5 built for "autonomous agents." All of that is optimized for one thing: getting you to write less code yourself. That's the point, and for a lot of work it's a legitimate trade. But it means the industry has spent the last two years running an uncontrolled experiment on how much manual coding ability you can outsource before it costs you something, and the bill is starting to come due in specific, checkable ways.

Where the gap actually bites

Atrophy isn't a vague anxiety — it shows up in three concrete situations, and it's worth being honest with yourself about which ones you're exposed to:

  • Interviews. Most technical interviews, take-homes aside, still restrict or ban AI assistance. If your day-to-day fluency has quietly migrated into "I can direct a model to write this" rather than "I can write this," a whiteboard or a live-coding round will expose the gap in real time, in front of the person deciding whether to hire you.
  • Outages. When production is down, the fastest path to a fix is usually reading unfamiliar code under pressure and reasoning about it directly — not composing a good prompt about a system you don't have time to explain. Incident response rewards exactly the skills (fast code reading, mental debugging, knowing the API surface from memory) that get least exercised when a model handles them for you day to day.
  • No connectivity, no assistant. A flight, a client site with a locked-down network, a tool outage on the model provider's end — any of these puts you back to unaided coding for a stretch. It doesn't have to be permanent to matter; it just has to happen on a day you needed to ship something.

None of these are hypothetical edge cases dreamed up to sell a product. They're the ordinary conditions of a working career. The question is whether you'd notice the erosion before one of them puts you on the spot.

Enter the drilling tools

That's the gap a new tool called Atrophy CLI is aiming at, according to a write-up in The Register this month. The pitch is blunt: treat coding ability like a chess Elo rating rather than a binary "can code / can't code." You start with a baseline exam that establishes where you currently stand, then the tool drills you across a handful of distinct skill categories — syntax recall, debugging, code reading, API memory, and problem decomposition — and tracks your rating over time, the same way a chess engine tracks whether you're actually getting better or just feel like you are. The tagline is the whole thesis in one line: "If AI assistance is quietly eroding your ability to code unaided, the chart shows you – before an interview, an outage, or a day without wifi does."

The design choice worth noting is the breakdown into separate categories rather than one composite score. Vibe coding doesn't erode all coding skills evenly — you can stay sharp on decomposition (breaking a problem into steps) because that's still mostly your job even with an AI assistant, while your syntax recall and raw debugging-without-a-copilot muscle quietly go soft, because those are exactly the parts you've been delegating. A single score would hide that. A per-category rating at least tells you which specific muscle to retrain, which is a more useful signal than a general sense of rustiness.

A necessary hedge

Worth saying plainly: Atrophy CLI is new, and I haven't seen independent data on whether its ratings actually predict interview or incident performance, or whether regular drilling meaningfully reverses skill decay versus just making you good at the drills themselves — that's a real risk with any practice tool that isn't the actual task. Treat it as a diagnostic worth trying, not a proven fix. The category it's carving out — measuring AI-assisted skill decay, separate from AI-assisted productivity — is the useful idea here, whether or not this particular tool turns out to be the one that sticks.

What to do about it without giving up the productivity

You don't need to swear off AI-assisted coding to protect yourself here; that would be giving up a real productivity gain to hedge against a risk you can manage more cheaply. A few concrete habits do most of the work:

  • Periodically solve something you'd normally hand to the model — a bug, a small function, a data transform — entirely unaided, and time yourself. Notice where you hesitate.
  • Before an interview loop or an on-call rotation, deliberately do a stretch of AI-off coding, the same way you'd cram before an exam you haven't needed the material for in months.
  • When you do use AI assistance, read the generated code closely enough to explain it to someone else, rather than pasting and moving on — that's the step most responsible for whether the exposure builds your understanding or just launders around it.
  • If a rating-style tool like Atrophy CLI works for you, use it as a smoke detector, not a training regimen in itself — check it periodically, and if a category is dropping, go do real unaided work in that area rather than drilling flashcards for it.

The honest framing is that AI-assisted coding and unaided coding ability are now two separate skills you're managing, not one skill with a shortcut. Employers, interview panels, and 3 a.m. incident channels haven't fully caught up to that yet, but they will notice which candidates kept both skills alive.