gail: An AI-powered GitHub Issue Labeler
8 points by atgreen
8 points by atgreen
I think @simonw once said “the best uses of LLMs are turning text into less text”, and this is a good example!
Question: it seems like “dry run” would be inaccurate because llms are non-deterministic. What if it instead wrote the tags to a json, and a separate script could apply the json tags to the issue?
Hi – I wrote this. When I started, I had two programs, as you suggested. However, the labels it assigned were the same on every run, so I merged the programs for ease of use. Maybe this won’t always be the case, but it was for every repo I tested with.
Here’s the prompt it uses: https://github.com/atgreen/gail/blob/6a2249752da5a97b87586099b0168ef1918ecc36/gail.lisp#L82-L86
Examine the following JSON object representing a GitHub issue. Respond with the Issue Number followed by a space-separated list of labels that are appropriate for this issue. Here are your label choices: … If none of the label choices are appropriate for the issue, don’t list any. Don’t say anything else.
I created a simple github action to tag all incoming new Issues using gail. It’s very easy to use: https://github.com/atgreen/gail-issue-labeler-action