Has the cost of building software dropped 90%?

I am shocked by the number of people who are dismissive of AI, or have stuck to the whole copy and paste into a chatbot approach to development.

I'm finding this stuff, when given proper guidance, can reproduce experiments I've run incredibly fast. Weeks of typing done in minutes of talking to Claude Code.

In the working world, a lot of the time what matters is getting results, not writing 'perfect' code the way software engineers would like to.

concerns: - security bugs - business logic errors that seem correct but are wrong

as long as you have domain experts, I suspect these will gradually go away. hopefully LLMs can be trained not to do insecure things in code.

> In the working world, a lot of the time what matters is getting results, not writing 'perfect' code the way software engineers would like to.

But you do recognize that one's ability to speedily implement features is dependent on the present quality of a codebase, right? Being a serious practitioner here means balancing active feature development with active tending to the codebase to keep it in a reasonable state, since its quality will tend ever downward otherwise.

In your experiments, do you find agents readily find this balance? I ask genuinely, I have only minimal experience with Cursor.

To be blunt and a bit nihilistic: I get paid to ship features.

Client wants a feature EoW, they get it EoW, they're not paying for a week of extra work for the "quality codebase" feature.

But the good thing is that we've had objective and automated tooling for quality checks for code. We used to use them for humans, but now we apply the same tools for AI.

Good unit testing practices, exhaustive linters, .editorconfig etc. force humans AND LLMs to produce code within specific parameters.

If your project doesn't have automated tests and linters, now is the time to add them. Maybe use an LLM to help you with it :)

Maintaining a quality code base is more than just having tests and linters. It's about organization, right-sized abstractions, architecture and choosing the right patterns. There is no real way to automate the verification of these things. If an agent farts out a 2,000 LOC feature in a day but bifurcates the code base, duplicates functions or makes awful abstractions, it WILL eventually turn into a big ball of mud.

All that being said, if wielded correctly an LLM can contribute to a healthy repository, but it requires much of the same thought and planning that development pre-LLMs did. I promise you, if you stick with the same code base long enough using your approach and little consideration to its health, it will become a hellish place to build in.