Remix.run Logo
▲ DanielHB 4 hours ago

I found that keeping my prompts small and incremental and reviewing the output at each step helps a lot both with keeping my skills sharp and keeping knowledge of the code. It is interesting that I actually learn new tricks from LLM generated code and keep it in check when it goes overengineering.

But that makes LLM speedup like 1.5x the manual-typing coding speed, vs 50-100x you can get if you vibe all the way.

I still use LLMs for troubleshooting, searching, codebase knowledge gathering, solution exploration, writing tests and documentation and whole-codebase changes and there the speedup is dramatic. However I don't feel I become a better developer for doing these kind of tasks without LLMs.

For documentation and tests though I feel I am slipping on reviewing the output because it is just so boring that I often "surrender" the task to the LLM which generates bloated tests and bloated docs.

▲pezgordo 3 hours ago | parent | next [-]

I divide projects in "care code" and "care result". I am buiding random tools all the time for which I dont check the code. For my long term project I do check the code.

This is the best I came with so far, but I have to admit how slow my main projects move, if at all.

▲JamesLeonis 2 hours ago | parent | prev | next [-]

I think I have the same experience, both good and bad. Smaller tighter feedback loops are like smaller PRs I review but at a much slower overall speed multiplier than straight vibing. Additionally I found I can use smaller context and even sometimes smaller models and get results back faster, but I must be more hands-on guiding things.

▲techmunky 3 hours ago | parent | prev | next [-]

There is still value IMO in creating atomic tickets for systems where humans are in the loop. Building a new UI experience? Have each component at minimum be it's own ticket. Maybe theres additional layers where you can break it down, scaffolding, skeletons, etc. For human led teams where AI is just making the code and helping with design decisions I think this is still the way to go for the bulk of the work. Can AI do everything soup to nuts? Sure but nobody is going to review that well and you can't revert part of it without taking out the whole change.

If you have a robust ticketing solution at your org just use it. Theres probably a baked in MCP. At work I use(d) JIRA (laid off, hmu w bay area or remote work), at home github issues are fine to a point then I use vikunja or openproject depending on the box its gonna live on.

If I was Junior in 2026 I would be breaking down work into the smallest components possible so that my understanding of the work was scoped and more importantly the inevitable AI sprawl would be similarly scoped (and easier to rein in/call out).

If you don't know how to break the work down ask for help. I wish I asked for help more earlier in my career.

▲pragma_x 3 hours ago | parent | prev [-]

I can attest to the first part. I think I've gone on record elsewhere in comments on HN that I much prefer using an LLM in a "sage" mode, where it functions more like a tutor/mentor than between me and the work. This has accelerated learning and, like you, I've learned new tricks this way.

As for the last part about bloat, I've started tinkering with prompts that attempt to minimize generated code output. I'd wager that having agents take a second or third pass to "prune" the output of the former would work too.