Seeing quite a few articles on "vibe coding" as of late. Meanwhile, in my own usage I find myself having a gut-level aversion to outsourcing too much to the language model.

Don't get me wrong – I use Claude pretty much daily as part of my workflow. Since the start of 2023 I've been happily leveraging the chat interfaces to various frontier and local LLMs.

I was able to add video uploads to this site in a matter of a day with Github Copilot. Language models are a profound workflow accelerator when you're working "in your lane", especially when extending a project with well defined pre-existing patterns.

locus of control

As I've leveraged these models more and more, I've started to have this creeping vague anxiety start to sink in. Part of what concerns me about leaning further into usage with agent flows is that I will no longer be able to retain a mental model of the project's architecture.

My current workflow looks something like this:

• Scope the ask to a specific "component": things like a request handler, a call to the DB, a React view / sub-view
• Include files in the context that "surround" the component and (if available) a file that contains methods of a similar structure
• Prompts are not particularly special, I generally just like to highlight the main goals of the task. For example:

Keeping my interactions with these models scoped as "conversations", I feel more in command and intentional with my interactions. I often times will not use the code output directly, but rather as a jumping off point to continue coding myself.

I guess my logic is this: forcing myself to read and adapt the output results in some neurons firing, my own cognitive processes are still at play. Every time I try to leverage agents, I've been getting uneasy about glossing over important details and shifting the locus of control towards the model.

Not trying to yuck anyone's yum, just reflecting on my own usage. I don't think I'm quite ready to fully immerse myself in the vibes.