Getting AI-fooled

Coding assistants are fantastic. I love using them and they’ve reignited my love for building software.

But they are not all-powerful.

When you’re working with AI it’s easy to get fooled into thinking it has all the answers.

It often doesn’t.

What’s worse, it makes things up.

The closer you are to its training set, the better the performance.

If you ask it to write a todo app, it can do it like magic. Boom. There it is.

If you ask it to implement something in a niche framework you need for your project, you’ll get something that’s worse than wrong.

It will be confidently wrong.

It makes up configuration values that don’t exists. It takes examples from frameworks similar to this one. It overrides implementations that already worked.

And if you just ask it to do better, it can’t.

But it will try to please you.

So, is it all bad? AI-code gen is shit, let’s move on.

Not exactly. You just need to be aware of the limitations.

The fix

For some things there’s still no substitution for actually reading the docs, looking at stackoverflow examples and finding blog posts.

I pipe the docs into ChatGPT 01, ask it to come up with a solution and write pseudocode.

I take the output, put it into Claude (Cline) which has my project context and it turns it into code that fits into my project.

This works like a charm, most of the time.

But sometimes, there’s still no substitution for going old school, rolled sleeves, straight into the docs, debugging the app flow, Sherlock Holms style.

Know when you hit the LLMs limits and break out or you’ll get as much bullshit as you ask for.

Yours,

Taj