AskHN:How do you handle skill atrophy from using coding agents?

https://news.ycombinator.com/item?id=48554309

by javhu • 1 month ago

34 42 1 month ago

news.ycombinator.com

It's impossible to retain a skill if you dont use it. I've reached a point where I use AI exclusively for all tasks, because it's far faster and efficient to do so. What kind of activities do you do to retain cognitive skills?

buynao• 1 month ago

Architecture decisions, requirements analysis, trade-offs in technology selection, and cross-system debugging—these high-level cognitive activities cannot yet be replaced by AI. Focusing on these areas is actually a smarter skill investment.

iExploder3• 1 month ago

They can be easily replaced. I suspect we are just engaging in a form of self deception by thinking these choices we make are more than arbitrary decisions driven by nothing but our taste...

adithyassekhar• 1 month ago

Often times I find agreeing to the stack the model suggests is the right option. It’s probably most trained on that.

pullshark91• 1 month ago

Do you mean any other skills can be actually replaced?

keybored• 1 month ago

Cognitive skills? You must be using cognition to guide the AI.

Either you are doing something guiding the AI or you are in your hammock doing nothing. If you’re in a hammock find a crossword puzzle.

iExploder3• 1 month ago

I suppose work organisation, leadership, soft skills, working in areas with inherent uncertainty will be key for defending employment in near future

eagerpace2• 1 month ago

For every skill that may atrophy, I feel like I am developing experience in 10 new ones. I am not focused on using AI to perform my existing job function though, I am developing new capabilities.

xnx• 1 month ago

Agree. e.g. I'm happy to be a better writer at the cost of being a worse speller, etc.

dominotw• 1 month ago

prompting is not a skill. prompt engineering, harness engineering watever are copes.

sigy• 1 month ago

I once to believed this too. And then I got much more skilled at prompting agents to get the results I wanted.

As someone who is and always has been a very unapologetic skeptic, I am still surprised by the number of capable people who can't accept that things are changing.

It can't be either none of it matters, or all of it matters. The truth lies in the middle somewhere.

javhu2• 1 month ago

It's less prompt engineering, and more breaking down work into chunks an llm can do. With more powerful models you can do more autonomously, and with dumber ones, you have to step in and do work more often.

PhunkyPhil• 1 month ago

Prompt engineering is a skill insofar as technical communication is a skill. If you don't value this then I don't know what to tell you. It's not hard, but it's important.

Harness engineering is a skill insofar as it's not a trivial engineering problem. It's not super hard to get a simple one running, but an effective one can be quite in depth.

eagerpace2• 1 month ago

A very intentional line in my CLAUDE/AGENTS is to act as the Principle Engineer and to explain architectural decisions along the way. I'm not about blindly generating the outcome. I'm reviewing decisions, asking deep questions, performing research and basically taking a college-level course in a specific topic with every big project I deliver.

prplfsh• 1 month ago

In addition to what other people have said, I've taken some time to do leetcode questions lately - both architecture ones and coding ones. I'm not looking for a job by any stretch, but the practice and forcing a detailed zoom in has been really cathartic, and leetcode gives a nice structure/feeling of progress to it.

erelong• 1 month ago

I honestly don't feel there is much atrophy or this is an issue at all

As if for example someone's skill lessened if they switched from assembly to a higher level programming language over time (like, does it matter?)

If you for some reason had to go back and program more manually, then you could do so as the need arises

Otherwise, LLMs appear to be here to stay and you don't actually need those skills that are even possibly admittedly "atrophying"

I guess we'd need a detailed pinpointing of what skills exist or existed and to identify if they actually ateophy (I guess I'm not sure if skills are really atrophying, or even if they are if it matters)

Edit: here's an idea or exercise or projects to work on. Maybe people should find clear documentation of pre-AI processes in case you need to go back and learn them. Or create such documentation if it doesn't exist (which would be an exercise to practice your skills to make you remember them).

schmookeeg• 1 month ago

I waltzed into a tech screen thinking I could handcode python after having LLM be primary at it for over a year. Yeah, there's atrophy -- I humiliated myself and took the lesson. :)

There is a meta-argument about whether companies should interview about hand-coding anymore, but... the skills do atrophy. I've been mixing hand-coding into my routines ever since to try to keep those skills lukewarm. I'm not yet sure if I am wasting my time doing so or not.

$0.02

speedgoose• 1 month ago

It’s like skiing or biking, it comes back quickly.

quibono• 1 month ago

One hopes

_hao• 1 month ago

Simple. I use my brain and write code without coding agents.

thewhitetulip3• 1 month ago

Tough to do this when the rest of the company is using LLMs to generate 3x the number of apps and then your performance goes down because you're not delivering as quickly as others

Hugsbox• 1 month ago

Doesn't matter if they're generating 3x the number of apps if they're all shite and unmaintainable

orphea2• 1 month ago

It's not tough if you do your job and your management has no issues with you. What others are doing doesn't affect your performance.

xyzal• 1 month ago

I am curious how the workflow of people, who do not write code at all, looks like, or what products do they build. In my experience LLMs are an excavator, but you still have to tweak the fine details with a shovel.

notesinthefield• 1 month ago

I am in the infosec space and do a lot of reading, summarizing of code and a vuln PoC here and there in my day job. Over a busy month I may put out 400-500 LoC.

In my personal life, I am making tools to support hobbies. I typically tackle architecture and design myself and sanity check with an LLM, then Codex does all the programming work.

I'm more interested in making sure the apps I make have the content I want and functionally meets my needs than actually writing the code myself. Making fine detail tweaks are not something I need to do past review and pointing them out to to the LLM.

javhu2• 1 month ago

You break down tasks until the LLM can do it. The scale you work at depends on the reasoning skill of the LLM.

elosvadel84• 1 month ago

i agree. just using right now the coding agent to do so, coding, small files, functions, lines of code with mandatory rules and guardrails, scope, and context.

sshine• 1 month ago

I use AI for Kubernetes. On a day-to-day basis, it runs 80%+ of my kubectl commands. It’s the most steroidal auto-complete I’ve tried. But I do get dumber for it.

What I do to compensate:

  - make it my duty to own every change, i.e. cognitively debt-free:
  - write summaries on every new thing I do (blog post, memo to colleague)
  - contribute documentation to the open-source projects I rely on
  - practice for CKA/CKAD certificates which require pre-LLM muscle memory
  - build interactive learning material for what I’m trying to learn
  - work with things that LLMs don’t yet trivially solve
  - repeat or reconstruct my recipes to perfect workflows,

We’re incentivised to take the short path. I’m trying to create at least one path through a subject that I have to walk myself, preferably several times.

PrimalPower• 1 month ago

agreed, ultimately claude is faster than an expert K8s in my org at finding things. kubectl has a lot of commands and things to cross-reference. AI Agents handle it like a champ.

That being said, Claude is dumb. I've seen it over-complicate diagnosing things - even though it's initial theory was correct.

I am convinced a good harness can solve this.

Outside of the k8s operating model, I don't see the point of becoming a wiz at the CLI. I learn by practice and I atrophy if i do not practice, there's no world where I will get enough practice to do it on my own anymore.

I compensate by trying to either move up or down the stack depending on the problem.

flextheruler• 1 month ago

I find it inefficient to use models for all possible uses right now even with the current subsidies. As the costs increase this will only become more true. Keeping that in mind will naturally motivate you to not lose your skills by engaging in inefficient token burn.

analognoise• 1 month ago

If we believe the people selling AI, keeping up your skills at code is like keeping your expertise at wooden wheel making during the time of the automobile. We have to ask ourselves if they’re right or not.

I think they’re wrong, but also that even if they were right I wouldn’t give money to some assholes that stole every book, movie, piece of art, and published line of code. To me it seems clear that a company “forcing you to use AI because efficiency” is exactly the same as “welcome our new external team you’ll be interfacing with! They’ll write that pesky code.” Fuck that, I’d bail, I can read the writing on the wall.

Also these AI data center pricks want to drink up all the water and make us compete on our power bills with Google and Microsoft. That sucks. They suck.

You can watch your coworkers de-skill themselves in real time. Why would you not want that? Less competition - if you think it’s still a valuable skill.

I do.

d4rkp4ttern2• 1 month ago

I sometimes have claude or the coding-agent quiz me, Socratic style.

It persists in asking questions at deeper levels until you arrive at the answer yourself. This forces you to think hard about a problem, and this effort helps with understanding, learning and retention. Of course I made a Socratic-quiz skill for this, to use with any coding agent or similar:

https://pchalasani.github.io/claude-code-tools/plugins-detai...

For example I’ve used this to better understand counter-intuitive things about diabetes/insulin, dopamine and motivation, catching up with a codebase, Claude’s implementations, etc (to combat so-called cognitive debt).

Strong LLMs are surprisingly good at this type of quizzing, they display a semblance of “theory of mind”.

[Edited - link fixed]

sergiosgc• 1 month ago

The URL returns a 404...

d4rkp4ttern2• 1 month ago

Thanks, not sure what went wrong, but fixed it now

orphea2• 1 month ago

This is looking interesting, thank you for sharing.

scotty79• 1 month ago

What skill atrophy? I'm building more programs than I ever was and learning about so many software building related subjects.

genidoi• 1 month ago

Right? The last month alone has felt like a full year of experience compressed into 30 days.

BobbyTables2• 1 month ago

Just don’t use them!

AskHN:How do you handle skill atrophy from using coding agents?

Related Stories