Practice Mode

This new "practice" has been very helpful for me recently, /practice-mode

In practice mode, I ask AI to write the code (or complete a task), but to do it in practice mode. In this mode, it has specific instructions like, dont deploy to prod, test it on stage, make a backup of the database, dont make changes if workspace is dirty, etc.

Then with all these guardrails, write a PR as best as you can, and record your learnings.

Three things come out of doing practice mode

• There is an opportunity to focus on learning. By preparing a PR, the AI discovers dependencies, discovers assumptions, discovers blockers and more. Some things are not apparent in a code scan, and the practice mode really pushes the understanding of the system.
• The guardrails encourages more to be done autonomously. Have a meeting, come up with 5 things to "try out". Ask AI to try it out one by one. You dont need to babysit it. You dont need to get it right. You are looking for an easy way to get more autonomous time from the AI. It also untangles prioritization, things that are not critical can still be run in practice mode. You only need to prioritize it when you bring it in.
• Review quality goes up. I can ask fancier models to review the PR, the task, and the learnings all together. I can ask them to review a set - and flag comparisons, rank by confidence, etc. When I review things manually, I'm also paying attention to the right details. I dont have to wonder as much - will this actually work?

My personal practice mode has gotten more rich. It takes screenshots, snaps DB schema changes, reports on evals, summarizes PR changes, etc. The goal here is to make it easier to go from practice run slop to ready for production.

The dual mode of "execute" and "practice" has been helpful for me - be careful though, you might end up with 50 practiced items and get overwhelmed trying to ship them.

Beginning again,
Ibrahim