Musings on LLM-assisted coding

I consider myself a "indie" developer ever since I've decided to take on Livtet more seriously. It's become both a project of passion and something to sharpen my learnings as I continue to look for work. I think anyone can consider this label as they work on things for the public that they intend to have some sort of audience consume or take on. But it's not just being a "software developer" anymore. Thanks to the advent of LLMs, folks can do what used to take days in hours^[1]. Speed used to be a defining factor of how folks in this speed produced applications. Yes, taste, design and affects all matter — however if your competitor or rival was able to release more features and design updates quicker than you can, your analytics can show you what's become apparent: users can and will drop off^[2].

This is to say, skilled development is possible and has been possible with LLMs for some time. I'm realizing that this is a bit of a controversial thing to say. Because of a near-allergic reaction folks might have with the technology that they already have been using to some capacity (be it through Google's search through the company's own admission), no one is actually trying:

• to see if there's options beyond the Big Three
• how these things actually work
• what and why frontier models operate

This works in favor of the companies that folks are allegedly against because it allows them to solely control the narrative on how these systems work. I do not believe that they'll be capable of replacing entire engineering teams but I have a bias as an engineer to believe that my work - knowledge work - is unique and cemented enough that nothing can surpass it. At best, we'll see (and already have) more labor reductions as more work is transitioned from conventional software engineering into what Laurie Voss called product engineering. If you can have a machine spit out robust enough code - which is what we already expected from mid-level developers and below to produce - then we can have the more talented and experienced developers shave off rough edges in an automated fashion and focus on business (or non-business, I'm not a MBA) value.

What does that mean for indies?

This is either a golden period for independent developers or a bleak point, depending on how you choose to operate.

To be a AI vegan is going to require a dramatic upheaval of how one operates. You'll have to forego using Windows, Linux or macOS since each of these operating systems deploy code generated by LLMs^[3]. You'll have to find new hosting providers that don't provide any sort of AI compute. You could be more of a ovo-vegetarian and stick to places like DigitalOcean or Cloudflare while running your human-grade software.

If you're more reasonable (or at the very least are doing the same thing when React took over the Web, or when Amazon's approach to Web development infected every aspect of the industry), you're probably looking at tools that aren't run by the Big Three (Cursor, Claude Code or Codex) because you want more control over how your harness works and are looking for something akin to how a game developer writes their own map editing software over time (except that you're mainly adjusting configuration and what models you're making use of).

Wait, so Jacky, are you—

Yes. I use LLMs daily for my software development. I've already written about my initial angst and then acceptance about them on here. They're not capable of doing anything by themselves because they're call and response machines but folks are working to make them more automated and networked.

In my own use case, I'm having it run across pages on https://livtet.olamaelcu.net/ and monitor reporting logs from Sentry to find bugs, file them on GitHub and have other agents do the work of picking them up and closing them with evidence videos and screenshots to prevent regressions. This would have normally been a few days' of effort for a gnarly bug. Now it's automated away in a few hours. And for someone like me who's trying to build something that can compete with https://fable.co/ or https://margins.app/, I'm willing to take on all of the help I can get.

All in all, I've accepted this new wave of development. The same way I accepted:

• React as the defacto means of frontend framework development despite being not a fan
• English as the primary language of software development despite not only speaking it
• 12 Factor development, which was a decent and good idea actually
• Datadog/Cloudflare over Sentry for application telemetry (Sentry's open enough to run at home)

I believe a level of knowledge work crisis of identity is happening for software developers and it's only a matter of time before we actually reckon with it. And when folks are ready, there's a few books^[4] folks should sit down and study before making another move.