AI-Driven Development (AIDD) Assistant: uses grokker as a backend to support a development cycle that looks like this:
- Write test cases, TDD style. Add copious comments to the test cases, as these will be read by the LMM backend to help guide the development process. These comments are where you express your intent about how the code should be written and behave beyond what can be explicitly tested -- shades of BDD.
- Run aidda. The tool will iteratively run the test cases, and use the test case output along with the comments to generate code that passes the test cases while converging on the intent expressed in the comments. When the test cases pass, the tool will generate some recommendations for further development and then exit.
See the output of aidda -h for usage for now.
I use aidda every day. It's a great way to at least get a first draft of code written and explore solutions to a problem space. I tend to step into the code (still using github copilot) to work out details or to get the algo unstuck when it writes itself into a corner. I've found that the 128k-token version of GPT-4 is pretty capable otherwise -- this won't work well with smaller token limits or earlier versions of GPT.
One interesting thing about aidda is that it also does a pretty good job of discovering and describing where the test cases themselves are weak, ambiguous, or just plain wrong; this has always been a pitfall with pure TDD or BDD.
At this point the tool is simply a shell script and is specific to generating Go code on Linux. I expect after the dust settles I'll likely use aidda.sh to generate aidda.go for a compiled version of the tool, and that will better enable the complexity needed to support more capabilities. I'm open to pull requests, but am otherwise pretty head-down actually using aidda as-is to work on some Go-specific projects at the moment.
-- Steve