This article shares why a web engineer holding a Claude Code Max 20x plan has mostly stopped using AI for coding at work. It's not about abandoning AI — it's about narrowing down where I let AI take over.
Why I stopped having AI write my code
My growth pace slowed down
Looking back honestly, I was acquiring knowledge faster before I delegated most coding to AI. Once I drastically cut down on writing code by hand, I quietly gave up a side benefit: the investment in growth that came along for the ride with the act of writing.
In learning theory terms: output (writing) drives retention more strongly than input (reading), as evidenced by the Testing Effect — the cognitive psychology finding that recalling material through output strengthens long-term memory. By losing those writing opportunities, I lost a quietly efficient learning channel.
I became less able to deliver quickly
Without the act of writing, the in-brain cache of implementation details fades, and I more often can't answer questions on the spot.
In practical terms: when a non-dev asks "can we do this?", my answer is more often "let me check" than before. From the non-dev's side, "well, in that wait time I'll just ask the AI myself" becomes a perfectly rational move.
The same pattern shows up in implementation and in conversations with other devs: "wait, how did we do this elsewhere?" now requires actual investigation rather than instant recall.
Sometimes it's faster for a human to write than for AI to write + review
When the implementation pattern is essentially copy-paste-then-tweak, asking AI to do it means I have to verify line by line (with diff) that only the intended spots changed. By contrast, when I copy by hand and then substitute, the work process itself guarantees "what I changed vs. what I left alone." The latter is often the faster path.
So what do I still use AI for?
Work outside my specialty
In domains where I have almost no background, AI can produce a passable, "good-enough-looking" output quickly.
Recently I tried to boost search traffic to a side-project web app and ran SEO measures — with essentially zero SEO background, AI gave me something "SEO-shaped" to ship.
Whether it's actually any good is honestly unclear, but once something is shipped a PDCA loop kicks in and things tend to course-correct toward "appropriate" over time.
Research
I'd rather hit primary technical sources directly than have AI summarize them. But for breadth-first surveys — collecting peripheral information, summarizing trends, listing options — AI is overwhelmingly faster.
People worry about "AI missing data it can't access (X, etc.)" and about "hallucinations." But web-information-gathering already assumes "don't take any single source at face value," so the probability of hitting a bad source via AI ends up as a rounding error, in my personal experience.
Recurring routine work
Things like running the app for manual verification, triaging Dependabot PRs, writing PR descriptions — I've increasingly delegated these to my own custom skills.
For example, instead of "spin up the container, seed the DB, hit it with Postman…", I let AI do all of that and just emit a curl command against localhost:3000. My job is reduced to "paste the curl command and observe the response."
The same curl trail also goes into the verification section of the PR description, and it works well enough. A nice side benefit: because every step is CLI, the reviewer can rerun the verification on their own machine more easily.
My personal heuristic: tasks that are "easy to articulate as a decision procedure and that recur frequently" are good candidates for skill-ification, and they pay future-me forward.
Local-only tools I'm the only user of
As long as I clear the minimum security bar, sloppy implementation is fine here.
I've personally built a few ruby-lsp addon gems for myself, which enable VS Code go-to-definition for Rails DSL methods (scope, delegate, etc.) and CodeLens display of model associations. The vibe is "as long as it runs" — and I haven't written a single line of that code by hand.
These use cases tend to fill up my Claude Code usage anyway, so I'm keeping the Max 20x plan.
Closing thoughts
Things feel settled at the moment, but technology moves so fast that I might be saying the opposite tomorrow. Personally I expect that "coding" itself will eventually get fully abstracted away by AI — and this article, which kind of swims against that current, may quietly disappear one day.
Even so, I'll keep enjoying day-to-day engineering through this turbulent era.
Top comments (0)