About Prompt Requests

Stumbled upon an interesting discussion on Twitter. Context: open source contributions. Key idea: contributors should bring not pull requests with code, but prompts by which the maintainer can implement their wishes with an agent themselves.

The idea is rich. I'm interested in reasoning about it. It seems this is also relevant not only for the world of neural networks and open source.

In the replies they criticize. They say we've gone too far. We trust models more than a random contributor. That this would only work if the model always produced identical results. Or that a neural network on your side can also make a summary (prompt) from the code. And in general all this looks like micromanagement when the maintainer wants to control absolutely everything. These are all fairly reasonable comments.

But the original idea is that a large piece of work, especially in the form of code changes, is hard to review and validate. A relatively small prompt in human language is much easier for the maintainer to read, and they can make corrections at an early stage of development. They want to see the contributor's flow of thought, which is harder to see in code.

And I like this idea for two reasons. First, it's always been this way. Ideally you (as a contributor) first think, plan, approve the plan with the maintainer, then get to work. Second, in the AI era, clear ideas matter more from people to people than code. Even outside the work context. Saw an interesting idea in some article, heard it at some meetup or in some podcast — went and generated a meta-prompt, got a POC and played with it, but in your own specifics. Liked it — went to finish and merge.

I think that in in-house development the position should be somewhere in the middle. There's not enough time to plan everything, you still rely on merge gatekeepers and developers having roughly the same mindset. You only discuss certain key points that can't be covered by automated checks.

Using AI as a translator from low-level to high-level language and back is still a great idea. We give agents, due to their absolute lack of context, much more detailed instructions than we give people. If we take these instructions and show them, for example, as a request description or during planning, everything becomes clearer for everyone.

Agents teach us to work with people, yeah.