Hello Guys, Ive been using 3 AI OS and like 8 LLM Wikis and what always happens to me is that they get messed up. Either the file structure gets confused and changed by claude or the log gets too long. In the end the context window and the token usage gets why out of hand. If you have any expirence please help. Thanks for all the responses

WHEN BUILDING IN CLAUDE HOW DO YOU BUILT A SINGLE AI AGENT TO HAVE ALL THE OTHER AGENTS REPORT BACK TO THAT ONE?? IS THERE A PROMPT FOR THAT ? WHATS THE BEST ROUTE I CAN GET A FEW AGENTS BUILT BUT CANNOT GET THEM ALL IN ORDER EACH WITH A SEPERATE TASK, TO REPORT BACK TO THE MAIN ONE

If you use Claude Code then I am pretty sure you know what is CLAUDE md Your CLAUDE md is set up right But your token bill is still huge CLAUDE md is not lazy loaded It gets sent on every single message, every session, whether you needed it or not So the fix isn't adding more rules to it It's making the rules you already have actually cheap and enforce I added just 3 lines to mine and output tokens dropped noticeably ╰┈➤When editing code, output ONLY the changed lines or the single modified block — never reprint the whole file. ╰┈➤No setup guides, no recaps of what you changed, no explanations unless I explicitly ask. ╰┈➤Keep responses tight. If one line answers it, give one line That's it. No 2000 token essay of "best practices." Just rules that remove waste on every reply Second thing I keep seeing People running Opus for everything, even one line fixes and simple questions Opus is for the hard stuff. Architecture calls, big refactors Sonnet handles almost all day to day coding at a fraction of the cost Haiku is fine for quick lookups and small edits Match the model to the task Your CLAUDE md handles the output side Model choice handles the input side Do both and the difference in your bill is not small What's one CLAUDE md rule that actually changed how Claude behaves for you?

I built my own knowledge system that is a second brain for my work and life. Three separate knowledge bases. One for strategy, planning and personal usage, one for emails and work, one for all kinds of building. All connected through a graph index with Graphify. How it basically works Each knowledge base has its own folder structure (raw data, references, live state, logs). Every file gets indexed into a graph that maps connections between them. A graph query surfaces related files instantly, so my agent doesn't have to guess where something lives. Important decisions get their own file with context and reasoning so I can retrace them later. The whole thing auto-commits to a private GitHub repo every 6 hours. This made my agent stop digging through files or losing context. With this system, I can ask it "what did I decide about X" or "show me everything related to Y" and get accurate answers. What do you think? Tips or things you would do differently?

Hello everyone! 👋 Concluded my first DAY of the AIS challenge: building my first newsletter using Claude Code. Honestly never thought this was possible with so little coding input, I'm so happy. The WAT framework is amazing, I can use it for everything that I want to build now :) For sure there is a lot of room for improvement, but I'm very happy with the first result. Thanks for this amazing resources! #AISChallenge