I tried to use my screen B&W in a few different occasions, but couldn’t stand it. This is not for me.

writing tests are another thing they mostly do a good job.

haha i think so, a few days ago i had to add "do not change any other code, just write tests" :)

Additionally the environment you run this task is also important.

- vs-code openai models are super dumb and mostly useless. Even free version of ChatGPT gives much better results.

- Gemini-cli (2.5 pro) only produces garbage, while in vscode it's my go to model for complex tasks.

My experience has been:

1. Explain as many details as you can, almost like a product/feature spec.
2. Use better (aka expensive) models for complex tasks. eg: Gemini 2.5 Pro.

If you don't describe edge cases, or side effects most models won't produce anything meaningful.

Looks good. Can it also work as a distributed scheduler?

So I gave up on error monitoring and will rely on user feedback. Because these tools would require too much customisation for not leaking anything sensitive.

White page perfectly fits the political atmosphere. 😅

Even though I created the DUNS number through their service and I cannot pass through. Do they expect me to not use my legal entity name?

That looks very interesting. Thanks for sharing. Never looked into how highlighter libraries work. Tree-sitter looks interesting as well, will look into it.

Look how neat is this?

3. ChatGPT way: Use regex and highlight/format "matched ranges". -> Seems like the easiest but have doubts on flexibility of this approach.

I guess I will start with number 2. Let's see how it goes.

2. Just loop through the json data and format each key/value based on their roles or types. -> Easy, requires handling the indentation as well.

I guess I am going to have to get my hands dirty. I see 2 ways + 1 ChatGPT suggested.

1. Write a JSON tokenizer -> fun but will take longer, and might be overkill.