(this post was written by hand btw - because I despise AI Slop rhythm, emojis and structure sooooo much)
I was recently asked to give a short AI-Coding Workshop for another company, which is outside my usual work. That made me recollect my thoughts on the whole matter, because there is kind of a rift in the industry, and also in the developer community.
On the one side, we have principled opponents and skeptics (especially on Mastodon!), on the other side we have the proverbial Vibe-Coders who use Agents without any supervision. I wouldn’t comment on the latter, but want to address the issues the former group usually has.
1.) Coding-performance related: results are worthless/the code is crap/Time takes much longer in total. - I can only explain this issue with either:
- 1a.) They use inferior models or agents - Try Claude Code first, I also had abysmal performance when I tried out different tools like Codex and Deepseek in 2025. Claude Code is a completely different league. Also, you can, and should parallelize the tasks if possible. With the usual Max-Subscriptions you can run multiple agents on different projects (branch = worktree) at the same time. For me, it is an enormous boost, especially in the breadth of our development - Doing small QoL improvements, scripts, admin-interfaces, going the extra-mile to make a new feature rock-solid and - beautiful, is now the norm. On the other hand the “One-Shots” of even bigger projects and features get more impressive every day. Even if the AI provides only the first 90-95% of a feature, makes it for us much more interesting to fine-tune and going the extra mile. Knowing when to switch the mode is also important: YOU need to test it, review it, optimize it and make sure, the code fits in your overall architectural vision. But we really couldn’t care less about the boring boilerplate.
- 1b.) Claude works best for well-trained languages and frameworks. The more your stack aligns with stuff that other people are using (and is in the training material), the better it works in general. Home-grown PHP, or proprietary Java stack? Bad. Bog-Standard Rails app? Using Best-Practices for your stack? Works fine. So, people using it on a non-standard language or stack might have expectably worse results.
- 1c.) People work on completely different products in regards to surface area. If you have a very successful - finished - app, then each issue might only touch a very narrow, very deep aspect to improve the product by a tiny percent. Maybe your role in the company is also tied to that feature, like “I am doing fraud detection”. On the other hand you have start-ups, or even B2B products like ours - that have hundreds, if not 1000 of screens and CRUD and APIs and external APIs etc. and you usually work full-stack - Each developer usually works on completely different features every day. I am not surprised if people in the first category do not find AI helpful. But I promise you, it can really help the second kind of developers (me) to bring meaningful, faster, and even better results.
2.) Politics related - those issues I can relate to:
- 2a.) They don’t want to send their code to US companies. Great point! I would really prefer the Mistral coding agent (EU), or even self-hosted models, when they reach similar grade of performance as Claude at the moment (or even to a state like mid 2025 would be fine for me). Self-hosted is not economically possible at the moment, so I don’t see that as a real alternative. I would prefer to switch ASAP myself, and I can understand and relate to that reason. In our scenario, we work on remote developer machines and use only our code, and don’t generally use customer data in development to minimize any risk for data leakage.
- 2b.) Don’t want to support AI craziness bubble, ecological externalities, increased hardware prices etc. Also, totally relatable! But I fear, just reducing your Claude consumption is not even a drop in the bucket, as other people will happily generate useless videos, music, photos at the same time, Running Clawd-Bot 24/7, which I guess uses much much more of those tokens and energy at the same time. AI regulation is similar to reducing meat-consumption, changing to electrical cars/reducing vehicles, switching to renewables or heat-pumps - It’s hard to have any influence as an individual on that broad direction of society, while the mass of people and especially political actors do not follow. The AI-bubble cannot pop soon enough from my POV, but I suspect, Gemini and Claude might survive that fallout. I tried out Music, Image generation, but didn’t see any point. I despise all the “Linkedin” Emoji-rich engagement style with their “delves” and text rhythms. If I see that kind of text, I will downvote, mute, block the person. But I feel generating useful code, and potentially “better” the life of others is a more meaningful impact you might have with the tokens given.
3.) Other issues - Besides that, other issues I also see myself:
- 3a.) Total crush of the Junior developer pipeline and of the “Software programmer” job description. It’s really hard for juniors to get a new job. Even at our place, hiring juniors now is a risk we would not take - Why? Because a junior/entry level like intern - takes (much) more time from the senior people than they provide net contributions. Juniors also use AI to generate their work for them. Unmotivated juniors see no point in reading any of the code and just dump them as a code review to the Senior. If the junior is just a proxy between code generator and review, the role is superfluous. So, critical thinking and a product-mind is something that a Junior needs to demonstrate. Same goes for the Job description of “Software-Programmer” (lowest, coding monkey tier), people who take a spec/issue and produce code - Something an AI completely replaces. Highly motivated juniors can use AI on the other hand, to better themselves, read the output, use AI to let them explain the patterns and get up to speed fast. The pipeline of software people will dry out soon, as starting a career as a Software Developer is now undesirable; the only hope is, that this will relax the whole market and only people, that have a genuine interest in making things will flush to the surface.
- 3b.) Atrophy of coding skills. A fear a lot of people have, and I can also see, is that by using AI we as developers will lose our precious hardly-acquired coding skills. Just before AI saw the light, I was proud of my (N)vim skills, saw myself as an expert Ruby (Rails) developer with 1X years of experience. With Claude, I reflected on myself and found, that coding (usually) always only is a means to an end for me. Of course, there were times, where I enjoyed the flow of coding with a cool music, headphones, full-screen terminal and uninterrupted coding flow. That thing is gone - at least economically. But I also, even more so, enjoyed building and shipping things and have other people use and comment on my work. I still read the diffs that AI produces, I don’t blindly accept every plan claude presents, I don’t vibe-code-to-production. Claude functions the best when in the hand of a strong-minded individual with the whole product - and architecture - in mind. I can highly recommend checking out Justin Searl’s Full-breadth Developer (https://justin.searls.co/posts/full-breadth-developers/) which deep-dives into this role-change. While reading that article I found myself nodding a couple of times. By not coding every day we can naturally expect to not better our “coding skills”, and even lose total speed if thrown into an IDE without AI. I am not sure if we should introduce occasional training sessions, like Code-Katas, Advent of Code etc., as We the Senior developers are responsible for all code changes and need to review, understand and test changes and suggest better architectural decisions. Similarly, like navigating on Paper Maps was replaced with Just-Follow-The Arrow Google Maps I see there a great atrophy of navigational skills especially in people that never used paper maps before or needed to orient themselves without tools. But we can still review the “Plan” of Google Maps before deciding on a route, because we might have context or different taste than the “Best” suggested route.
Puh, those thoughts were lingering in my head for a couple of weeks, and I hope someone out there might see a value in their personal evaluation of AI in context of coding. I (and my colleagues probably agree) see it as a great enrichment of the work, but also in our company, we always nourished a leadership mentality, taking responsibility and great freedom of choice for every role. For us, AI resparked the flame to build cool new tools, admin-interfaces, small scripts, better frontend design and helped close a lot of old issues that we never had the bandwidth to tackle. And all of us now have personal side-projects that got so much more care than before.