i feel robbed of my own progress when i rely solely on AI to code for me. there has been, over the past 4 years since i started working as a software engineer, a clear increase in my skills as a programmer. at my latest new job - i am surrounded by people more talented & experienced than me, and i feel like i rob myself of the opportunity to actually learn & grow by trying to force AI to write the code for me, rather than do it myself.
i have aspirations, too many projects, that i felt the need to go to vibe coding as a crux. i was able to alt-tab between 4 different projects, just being a code reviewer & prompter, and i felt a similar flow state to original coding. but after a 2 month break and trying to get back into it (mycelia), im left with the consequences of my own haste.
the projects aren't laid out in the way i would in my head, so im searching for things that i thought would exist where they should - but don't. the code isn't of a high quality - it's verbose and over-fitted to edge cases & bugs that the ai managed to pickup, but has holes that it would've missed. ironically - it fills up it's own context window as if it gets a prize for writing large files.
part of what i've learnt at my current job (one of the faang ones) is the benefits of writing less. functional-style programming that favours composition of base building blocks in order to structure higher-level business logic with ease is something that i'll likely spend the next 10,000 hours of my career mastering. im robbing myself of that opportunity to ask "how could i write this better" when i don't go through the motions of typing that text out myself.
the cracked engineers aren't using ai to save time coding. they're using it to save time identifying areas in the code. i almost want a "read-only" mode for the ai. getting llm's on the device through things like copilot, and then cursor & now claude-cli and codex (and opencode, which i use) - have been a great advancement. i think giving these cli-based tools write access was a mistake.
ai is at it's most powerful when summarizing - that's why i started working on chamber, i always knew this. it isn't great at creation, it's not really generative, it's transformative. no wonder the discovery that kicked off the llm cold war was the transformer. i'll probably spend some time now investing in how to get the most out of a read-only AI. i can see a couple uses, but it has some pre-reqs:
- indexing a code project
- understanding how the project went from spec -> implementation (git timeline)
- updating with tips on how to code better after each commit (pr bot but in terminal, not on a UI somewhere in the cloud.)
- agent baked with knowledge of how i want to write code
- functional
- concise
- composition over inheritance
- multi-faceted
- favour easy testability (in-memory representations of all logic)
- what to work on next (ingests todo lists, project spec, and current progress)
let's see where this goes.
Top comments (0)