I recently made 2 skills called /initit and /outit (like a buffed version of the built in /init command in Claude code) which has solved that exact problem. Basically, the /outit command audits the code at the end of my session, creates a document of where we’re at, like a very detailed logging of the session then updates the claude .md file. The /initit is basically /init, but it prioritized the most recent audit. Naming conventions are how it’s all organized, chronologically so I can always go back in time. I find that it costs more tokens and takes more time at the start when I do this. Prompt 1: /initit and gain context of where we’re at Prompt 2: Audit the code base, particularly focusing on these areas. This is an audit, not an implementation plan, but it will be used later to create an implementation plan. Produce a document called ‘audit example doc’ which lays out your comprehensive understanding of the code base, particularly focused on these areas. Prompt 3: Create an implementation plan broken into smaller milestones to ensure that each milestone is achievable to achieve success. Success = add these features ‘list features’ without introducing new bugs. Prompt 4: /outit I’m not sure if that helps, but I hope it does! That’s the workflow that’s been working for me, I’m able to pick up where I left off because the Claude .md file is like a long term memory document while the audit is the short term memory from the last session, and together (create the postmortem audit right before updating the Claude .md file at the end of each session) facilitates a great handoff into the next idea, for the next iteration of the code. Good luck!

Hello! I'm experimenting with building a cutom harness too, like how Pi works! Using Antigravity 2.0 for the build. Thanks for sharing :)