-
@ 06639a38:655f8f71
2025-05-01 12:25:58
This took me more than an hour to accomplish this with Goose + Claude: https://github.com/nostrver-se/nostr-php/pull/88#issuecomment-2841986088
It's was really funny to see how the agent was going back and forth fixing errors. Some of those fixes where like 'Ah the depencencies needs to be updated, so let's run go mod tidy' while there where no dependencies added or whatsoever.
My feelings are still mixed how to use these so called AI tools.
-
@ fd208ee8:0fd927c1
2025-05-01 10:45:18
Yeah. 🤣 I was trying to think of how I'd make the resulting code really easy to read, but that puts all the complexity into the code-generation process. Woops.
-
@ fd208ee8:0fd927c1
2025-05-01 10:06:06
Instead of just bitching at you, here are some tips for getting code out of your AI that even a beginner might be able to read and understand, and that might not be complete garbage:
Include in your prompt a file of instructions like:
1. Do not make any function longer than 5 lines of code, excluding documentation and logging. Limit all files to 250 lines of code.
2. Make the code highly object-oriented (or highly functional, whichever you prefer or depending upon your use case) and modular.
3. Include a logger, integrate it into every function, and make it so that I can run the app with different logging levels.
4. Write extensive user documentation for everything, including type definitions and examples for implementation.
5. Use constants wherever applicable, and group the constants at the top of the class or module.
6. Use private or protected wherever useful.
7. Refactor and simply the code, and remove any dead or redundant code, after each pass.
8. Separate concerns, and give all functions a human-readable name.
9. Add a description of the architecture, patterns, protocols (like Nostr), and libraries you expect to be preferred.
10. If implementing a GUI or other user interface, point it to an example of one you like and tell it to give you something similar.
Then submit your request in Gherkin scenarios, instead of prose.
Reorganize the scenarios into features -- which should be relatively small, like "broadcast button" -- and get it all sort of comprehensive and coherent, and then feed the AI one feature at a time, along with the prompt instruction file.
And then work that one feature until it's good, manually test it and refactor the code, AND THEN COMMIT THE CODE TO GIT, before asking it to do the next feature. Start a new conversation before the next feature. This is like high-speed agile development and will ALWAYS get a better result because it's iterative and therefore more like how the best human developers work.
If you ask it to fix some code, open a different chat and tell it that another AI wrote the code. Even better, ask another AI and tell it the name of the other one. Just trust me on this one.
Once you have it doing what you want it to do, and the state can be fixed and tagged, have it write unit tests to freeze the functionality. Then make sure the unit tests are easy to read and understand. Remove any test you don't understand. Avoid mocking relays. Use a local or ephemeral relay for testing, instead.
And then write your own integration test, with some realistic examples.
-
@ 06639a38:655f8f71
2025-04-30 14:43:42
Agree, categories are better.
Topics are more related to conversations (a forum has topic for example) and could be categorized.