"Luxurious" Software Development with AI

If you ask any even slightly jaded software developer, they've likely worked on a project where having "nice things" cost too much developer time according to the folks in charge.

The first thing we were told was too costly, before Ruby and RSpec made specs cool, was a robust test suite. Others included speedy CI, E2E tests, full type definitions, code documentation, up-to-date guides, good READMEs, Storybook, Design Systems, Prettifying, Linting, CI, Deploy scripts, monitoring scripts, CLIs, the list goes on depending on your stack.

Or, if you had one of these after an enterprising developer spent some of their minimal free time getting it over the line, the feature would degrade if the whole team wasn't bought in and there wasn't time allocated to keep it current.

At larger companies, you'd get a few of these, but they'd be implemented more as straitjackets and somehow slow the whole process down rather than speed everything up.

AI-native software development with LLMs has inverted the math, however: these things, which felt like luxuries when developers wanted them, are significant performance and speed enhancements when coding with LLMs. Their value is visible in token usage and bottom lines. They pay for themselves quite visibly. Why is that?

LLMs really need two things to work effectively: they need to get up to speed quickly, and they need a feedback loop to keep them on track.

As any individual Agent or Subagent is born, lives, and dies in max a 40ish minute window (about the max time it takes a Claude Opus to fill up its context window for me), getting up to speed quickly without adding unneeded context is important. What helps with that? About half of those luxuries - code documentation, guides, READMEs, Storybook, Design Systems - all help. Making sure they stay up to date is an extra skill or two and a few extra lines of prompts. Unlike telling Frank that he forgot to update the documentation. Again.

Most of the other half: Speedy test suites & CI, E2E Tests, Type Definitions, Prettifying, Linting help the feedback loop, making it effective and fast when LLMs write their Mediocre Code.

The few of the remaining pieces - various scripts and a CLI - have been reduced in cost by so much with LLMs (and they'll even write tests for those, unlike me) that they're effectively free to write and keep up to date.

It's funny now that we're not personally writing code as much, development has become downright luxurious, but the good news is that these things also make it easier to dive into a new codebase with an LLM.

If your codebase is not quite luxurious yet, see the 3x3 Rule of AI-Native Software Engineering, especially the Constraint and Review parts, to help get you there.