I ran an experiment of automatically generating subtasks for ubiquity/pay.ubq.fi#421
I told Claude to provide a time estimate based on the issue specification and divide it by 15. So a 15-hour task yields 1 hour.
Then take a look at the existing Time: labels of the repository and label with the best fitting one.
Due to some problems with Claude this week (horrible at following instructions, rumored that they are quantizing the model to save on costs this last week) I switched over to Grok. Grok code seemed to not need the time offset.
Make a plug-in that listens to issue created or specification edited and automatically sets a time label based on similar instructions.
I do prefer using Claude because it's included with my Claude Code 20x Max plan.
Try getting it to work with claude -p "my prompt here"! I added a couple extra hours here cause of prompting and testing. I imagine if it only has to worry about a single issue, it should be better at following instructions.
Another remark is that I specifically tried to remove bias by telling it to ignore the original time estimate listed. Just provide a new estimate.
Make the offset configurable, because I imagine that as the models improve, this offset will change.
I ran an experiment of automatically generating subtasks for ubiquity/pay.ubq.fi#421
I told Claude to provide a time estimate based on the issue specification and divide it by 15. So a 15-hour task yields 1 hour.
Then take a look at the existing
Time:labels of the repository and label with the best fitting one.Due to some problems with Claude this week (horrible at following instructions, rumored that they are quantizing the model to save on costs this last week) I switched over to Grok. Grok code seemed to not need the time offset.
Make a plug-in that listens to issue created or specification edited and automatically sets a time label based on similar instructions.
I do prefer using Claude because it's included with my Claude Code 20x Max plan.
Try getting it to work with
claude -p "my prompt here"! I added a couple extra hours here cause of prompting and testing. I imagine if it only has to worry about a single issue, it should be better at following instructions.Another remark is that I specifically tried to remove bias by telling it to ignore the original time estimate listed. Just provide a new estimate.
Make the offset configurable, because I imagine that as the models improve, this offset will change.