IT/Software career thread: Invert binary trees for dollars.

ShakyJake said:

For those of you using AI coding agents for work, what's your typical workflow? Do you mostly stay inside the IDE, or are you primarily using the CLI or desktop tools like Cursor, Codex, and/or Claude Code?

Click to expand...

We use Cursor. Right now I am designing AI Agent standards for the org to use as I've done it the most. Not to say that's a lot but my past year or two of doing this is a lot more than just about anyone in this company. Cursor has a CLI offering like Claude now but for me personally I have used IDEs for over 20 years now. The user experience of just using it within what I know is fine. If Claude is somehow better it is eclipsed by the cost. Cursor is able to economically conserve token burn rate. Google's Antigravity is even better at this supposedly but I have not put much time to it.

The goal is to analyze all of the projects in our organization, design Cursor Rules/Skills/Hooks/etc to facilitate broad standardization. As this company grew immensely and its 10 times larger revenue wise than they were 4 years ago they have a lot of growing pain. Nothing is consistent at all and their primary projects are named around whatever their leadership at the time thought was cool. The primary application that runs the organization is named after a children's movie I shit you not.

This is my life right now:


All that being said token usage rate is going to be a big deal as companies seek to manage AI cost. Right now Claude is looking like the one you keep around if you're willing to really spend money on a few engineers but you do not want the org to use it as a whole. Getting a whole org into a different IDE from their usual one isn't as big a hurdle as you would think. But they are taking some time to understand why to have rules, skills, and what have you. They are all onboard with it though so I suspect we'll have a clean ship running here in like 6 months.

I have kept warning my superiors about this (big push toward Claude here and no one seems to be taking costs into their calculus atm) and keep getting hand waved. Our mostly non-technical leadership seems to think this is a magic wand that will cut staff costs and solve everything, even though some of us have demonstrated the impacts of data poisoning and feedback loops in the LLMs, plus modeled what long term costs will look like. This is the same company that tried to replace everyone below management with Poos at one point and we are still recovering from that fucking disaster. One would think that with the track record of our leadership fucking things up and the massive amount of cash that cloud computing is sucking out of us that someone might think twice about just grabbing ankles for this shit, but nope. No structure and no control, just like everything else out company has done.

Khane said:

Token usage isn't going to be a big deal. It's going to be the ENTIRE deal.

Claude cowork isn't bad with token usage. Claude Code is an atrocity with token usage. That being said I honestly feel that companies should only let very senior engineers have access to claude code subscriptions and it should be used sparingly for mundane, but low token usage tasks, like refactoring small projects, scaffolding new projects, building out test harnesses and cases, things like that.

I have a claude subscription now and have been playing around with claude code and while its very good, its somewhat slow and it will, very literally, make you a worse software engineer over time if you rely too heavily on it. A junior dev using claude code will never learn a fucking thing.

Click to expand...

We've just frozen all hiring above junior level due to AI first policies and w're even backing out of that to grab interns. VPs believe that interns + Copilot is going to invalidate experience and seniority. There's no reasoning with it at this point so I'm trying to mitigate by putting strict training wheels on how much AI can fuck my green/intern devs code by including documentation and requirements into rehydration, markdown files within my repos in individual components and classes so at least the LLM or agent has footing to make their terrible code decisions. My job has turned into vibe coding janitor and there are way more interns -> mids with Copilot access than there are Me so it's a never ending, fruitless race to fix what gets broken everyday.

Me and some friends who all work at different places have just conceded to the fact that whomever is senior and above right now is basically it for the foreseeable future for competency and even seniors are seeing diminishing skill levels by overuse of AI so our plan is to just ride this shit out as vibe code janitors and hopefully eke past the finish line as the last generation of actual developers.

The confidence of some people in committing the most retarded code imaginable because AI signed off on it is quite remarkable

ShakyJake said:

Where are they getting this from? Because in my org, pretty much everyone is clueless. Sure, we have access to Github Copilot, but that's pretty much it... and with zero guidance on how to use it effectively.

Click to expand...

I have no idea sir, I think they are just hoping that it works with no backing data other than cost analysis. We've also quit hiring specific disciplines and are all full-stack so I'm fixing absolute dogshit AI react code while still having to explain to people how to make paginated API endpoints so we aren't in a place to even do basic, non-AI infected code engineering right now.

I tried to warn against the same thing that Khane brought up, that really only seniors and up should be actively using this and event then sparingly and with great criticality for what they are given but management seems to think the opposite. It's a senior engineer in everybody's pocket. We're revolutionizing the tech space by making engineers irrelevant and spiraling into a massive year of tech debt and recovery once that proves to be a fiction.