My main point is that we need to expertly drive these tools. I forgot the trivial instruction and ended up with something that more closely resembles modern software instead of what I personally value. AI still requires our expertise to guide it. I'm not sure if that will be the case in a year, but it is today.

You seem intelligent so it is probably confusing to many why you are posting this.

You call it a trivial instruction, but it is not trivial. It was a core requirement for your own design that you neglected to specify. This is not different than leaving out any other core requirement for a engineering specification.

Most people would NOT want this requirement. Meaning most people wouldn't care if there are package dependencies are not, so the agent 100% did the right thing.

Exactly. The vast majority of people prefer to re-use code that has been battle-tested by use in thousands of other projects, where bugs have been found and fixed.

This isn't to say that they're perfect or bug-free, but they generally don't just save time but increase the quality of your final product.

Writing your own implementation of third-party functionality that already exists has its place, but it's generally something you implement later as an optimization for very specific reasons.

Right, and this is the core lesson: constraints have to be explicit every time, or they don't exist. The model has no persistent memory of your preferences across sessions — you have to re-declare them.

The pattern that helps: put your invariant constraints (zero-framework, no external deps, specific language version, etc.) in a structured template that gets included with every prompt. CLAUDE.md does this at the project level for Claude Code. For one-off prompts it requires more discipline — which is exactly the gap I built flompt (https://github.com/Nyrok/flompt) to address. Visual constraint block you can reuse across prompts.