The tradeoff between human and AI context

I’m working on a package for Laravel (iambateman/speedrun) that enforces a workflow to tell Claude Code how I want it to interact with my codebase.

The theory is that it should set up a staging folder of recommended changes first, and then implement them.

There’s a lot of interesting work to be done around making it safer to let Claude code make big adjustments without going off the rails.

So far, I’m noticing that… (1) it sometimes secretly chooses laziness…it leaves notes like “to implement later.”

(2) or it tends to do too much…it wants to go off and work on unrelated parts of the app even when not asked to. This is the most problematic.

(3) It tends to write tests that are hyper specific, and look for exact strings that will likely change over time.

As long as the applications are being used by humans and paid for by humans, we will need some group of developers providing context to the machine about what to do. Right now the tooling is outpacing our sensibilities about how to use it effectively.

The title of the blog post is currently "The tradeoff between AI and human context" which can have a subtly different meaning to the HN title