Ask HN: How are you preserving your skills while using AI? | Hacker News Reader

Ask HN: How are you preserving your skills while using AI?

9 pointsrdrmc8 comments

I'm a senior engineer at [Big Company], and AI tools are ever-present. There's no mandate that you need to use them, but they are so readily available that most people do anyways.

There's a lot of society level concerns with AI, but on a personal level, I'm starting to slowly feel less skilled than I used to be. I can certainly do more, but I understand less. The "Prompt-Then-Review" loop of coding harnesses (Claude Code, Codex, Pi, OpenCode, Amp, etc.) just simply do not encourage mastery in the same way as shaping the code yourself. Sure, you can argue you're "thinking at a higher abstraction". But what happens when that abstraction fails? as abstractions often do.

It's not a fast process this skill erasure. I'm not magically losing my ability to code overnight. However, it feels like rust. Slowly eroding the pillars until they give.

This tool (currently) needs a skilled hand to guide correctly. However using the tool feels like it degrades the skilled hand. This negative feedback loop I find truly concerning from both the ability to make a living in software and the ramifications on software quality writ large.

So, I ask HN. How is the community protecting their skills? especially when actively using AI.

I’m in this exercise right now, and I’m thinking about, how do I stay in the drivers seat. I think it’s a version of, bionics not robots.

Im trying to go back in time to when AI was more autocomplete than agent and write my code with assistance. Find the file, find where the code goes autocomplete the fix. Rather than one shot multi file changes do one at a time.

The bosses probably don’t want people spending 4 hours figuring out why upgrading a package conflicted with another package in weird ways or changed the API or other obscure things, so like, give troubleshooting 10 minutes, try to grok the code context, then prompt for a fix. Or prompt “show me where I need to look to fix this” and fix it yourself.

The question to really ask is what time scales are we talking about. Sitting down to do a task getting context and prompting code probably takes like what an hour, maybe 2 if you have to review it or fix something? If you add another hour to that, and really understand it, I think you’re not hampering your productivity all that much if at all in the aggregate.

The other thing I’ve always done is let myself get nerdsniped easily. Maybe multiple times a day if someone posts an error in chat I’ll give myself 5-10 minutes to look at logs/graphs/google. Keeps the job interesting but also I think keeps my mind flexible and into the details across a broad swath of tech.

That's probably one of the most sustainable approaches. Queries like show me where to look for the problem often turn out to be more useful than just fix it for me, because they leave the most valuable part of the work to the engineer

I'm facing the same problem. I'm spending a lot of time on side projects now precisely because my company is forcing our strongly encouraging everyone to use AI for coding. At some point I found myself complacent enough to accept AI code at face value which scared me. With my side projects, I still use AI, but only when I get stuck, to help me get the aha moment. But most of the time it's, hmm... That works but, I don't think we should do it this way. Very tempting to continue prompting and keep going from there though, I have to say.

I had a similar experience with unfamiliar frameworks. The less you understand the domain, the easier it is to accept clean-looking but wrong code

Good test actually : try explaining the solution without any hints. If you can't walk through why the code is structured that way, then the review was superficial

Ooh that's good advice! Very useful for everyone in my team. Appreciate it!

> But what happens when that abstraction fails? as abstractions often do

One of the upsides to agents is that you can try out several abstractions instead of just one. You should absolutely be reading / spot checking the output (depending on context).

While skill loss is often the focus, there is also skill gain. I have learned new coding patterns or framework features through these tools.

On you main question, effort. Like everything, you have to use it or you lose it. I never saw my skill as a developer being the part where I peck at keys. It's the thinking and designing process, which I get to do way more of now that I don't have to peck.

Recommendation: build greenfield side projects, exercise building out all the parts. This will make you better at both being a developer and an agent manager.

I just don’t use any AI tooling for my home projects. Work pays for results so AI is basically mandatory to meet the deadlines - but home is for fun so I can take longer to produce hand crafted code.

coming from a different pov, I was (am) systems engineer who never became a proper coder, then founder, now mostly product/growrth. So AI and agents dint erode a mastery i had, they gave me one i never had. What you describe though makes sense for the skills I did have and I see a dependency as well. Things I used to work out myself, even slowly, I now reach for the agent first by reflex. The speed effect is indeed there. I get anxiety many times that I don't understand subject matters or I learn them slowly. Haven't found a good way other than keeping some of the writing or critical judgment calls to me. On a happy note i try not to push LLMs at all to my kids as I start reading such dependencies happening there.

Reply on news.ycombinator.com