Peter Steinberger posted a compact prompt that turns tools like Cursor and Claude into persistent refactoring partners, directing them to keep iterating on architecture until it feels solid while enforcing live tests after every meaningful change and automatic commits only for passing work.
@MatthewBerman /goal refactor until you are happy with the architecture. ensure you live test after each significant step and autoreview/commit. track progress in /tmp/refactor-{projectname}.md
I use Cursor, Codex, and Claude Code. I want to make sure I use all my tokens during each period. What are some recurring tasks I should do on my codebase?
The shared instruction adds autoreview before landing code, progress tracking in a temp markdown file, and preference for higher model settings plus existing agent guidelines for verification.
Steinberger suggests discussing direction first on larger projects, yet no adoption numbers or head-to-head results appear in the post itself.
Positive users praise the shared AI workflow for iterative codebase refactoring as useful and worth trying, while negative users criticize it for high token costs and risks like flawed self-verification loops.
@MatthewBerman probs makes sense to first discuss direction and steer, but if it’s a smaller project even this yolo prompt will yield good results and deslop. ensure you use my autoreview skill
@MatthewBerman /goal refactor until you are happy with the architecture. ensure you live test after each significant step and autoreview/commit. track progress in /tmp/refactor-{projectname}.md
@MatthewBerman live test should include computer use, browser use and/or keys and crabbox whatever is needed to fully e2e verify. You want to find bugs during refactor not introduce. and use high or xhigh.
That stuff I have in agents md since this should always be a step before you land code.
@MatthewBerman probs makes sense to first discuss direction and steer, but if it’s a smaller project even this yolo prompt will yield good results and deslop. ensure you use my autoreview skill
@steipete @MatthewBerman i'd make /tmp/refactor earn its keep: 3 lines after every step - changed, broke, tested. otherwise the agent starts inventing architecture debt.
@chengyongru @elbomnetanel @steipete @MatthewBerman https://github.com/rbbydotdev/codemap-mcp
@benlatger @MatthewBerman “keep track of work in <file>”
@steipete @MatthewBerman Doesn’t reccuring refactorings just makes the models find false architectural problems and ‘fix’ them?
@steipete Trying this now
@MatthewBerman /goal refactor until you are happy with the architecture. ensure you live test after each significant step and autoreview/commit. track progress in /tmp/refactor-{projectname}.md
@elbomnetanel @steipete @MatthewBerman This is a good point, you can use ast analysis to find cyclomatic complexity for deeply nested ifs and Louvain to flag possible improperly split domains via module “network” topology
@rbbydotdev @elbomnetanel @steipete @MatthewBerman make sence
@steipete @MatthewBerman @bradmillscan
@steipete @MatthewBerman i felt the token burn just reading this
@steipete @MatthewBerman hot
@steipete @MatthewBerman 这是一串有神奇魔力的代码
@steipete @MatthewBerman This is like asking a student to grade their own assignment.
@steipete @MatthewBerman I will try it - thanks for sharing. Most people dont know all these secret commands
@steipete @MatthewBerman how do you put a limit on it?
@steipete @MatthewBerman Why is OpenClaw so sloppy if this works well?
@steipete @MatthewBerman This is a great idea. I'm always trying to burn my weekly subscription every Thursday and have to make stuff up.
@steipete @MatthewBerman There’s got to be a right answer in any loop is the issue
@steipete @MatthewBerman if you want to try something with a diverge-converge strategy (also a good way to standardise loop behaviour)