I pretty much use claude for everything at the moment. I'm a software developer, and to be honest I really enjoy the way sonnet (at the moment 4.6, 5 uses my tokens too quickly :) writes code. All of my routing and context files seem to work great with claude so no issue there. My issue is the more and more productive I am getting with the system the quicker I am ripping through tokens, so I have started to palm off different tasks to different models to see how successful I can be at still maintaining a smaller cost footprint with the two subscriptions I have at the moment (claude and chatgpt). I have just asked codex to update some of my context files and the edits have specifically called out codex by name... i.e.... ```Codex should not perform this process. Codex should tell the developer which step is needed and stop for the developer's result.``` (I like to be in control of when the database migrations happen for example). However these instructions are not for Codex alone, they are for, quite frankly, any AI or Human that reading my codebase. Okay, mostly AI, but the main reason for documenting all of this specifically is it forces me to document my process to allow anyone (AI or human) to pick up the project in the shortest amount of time. This is the reason I love ICM. I'm getting off topic, but how do I re-write this in a way that makes sense not only for codex, but for claude, gemini, kimi, human, whoever is reading this? Do I call out AI specifically? "AI should not perform this process"? Do I use inclusive language like "We should not perform this process"? (indicating it might be a BA doing this, as the next line will be "We should tell the developer...."? Quite curious to hear peoples opinions on this, trying to keep the documentation as documentation instead of specific to a model.

YOU GUYS 😭 😭 😭 This is torture on my soul and I think im going to end up divorced over this because I feel his desire to do, I really do, and his remorse, but Claude and I are just at a stalemate and neither of us know what to do. First off - LONG POST WARNING: These are two copies of my current claude sessions in VSCode Terminal. It just wont do as I want. And when the saying goes "happy wife, happy life" well.... you can only imagine how this has us (and by "us" I mean "me") in all sorts of a twisted up mess!! I genuinely don't know how this is possible but Im sharing a sampling of two of the messages I'm getting now almost every session. I've tried EVERYTHING = and before y'all say it, i've rewritten my Global claude.md, project claude.md file, config file, my memory.md (which i have since removed), AND SO SO MUCH MORE.etc. and these are still the messages I'm getting back (AM AT MY WHITS END): ─────────────────────────────────── ────── Example 1: Me to claude: Just tell me what you do wrong regarding the routing after you get a request. ──────────── 1 new message ───────────── I skip the router and navigate from memory. Your request named "the comms door file." The prescribed flow is: find the row in CONTEXT.md, go to that room, open its door, then work. What I actually did was ls business-os/comms/ and open the door in the same breath — because I already knew the path from the map and current-state. I never read the row. That cost real things, not just process points: - The comms row has a second column I never followed. It says Twilio/Retell via platform/. Every Retell hand I used is routed through there. I went straight to api.retellai.com instead. - The router has a credential rule sitting right under that table: "If a task needs vendor credentials, route to the broker boundary/docs first and ensure the automation is declared in platform/broker/manifest.json." I used broker credentials four times and never opened the broker boundary docs.

https://www.instagram.com/reel/DaYoXM5v8Xt/?igsh=ZHExa3NvOWx2MTJi

Ornith-1.0: the open-source coding LLM that matches Claude Opus 4.7 — and runs on your laptop. This open-weights agentic-coding model hits 77.5 on Terminal-Bench and 82.4 on SWE-bench Verified, beating Claude Opus on both, then ships a 9B size you can `ollama run` locally. MIT licensed, four sizes, and it taught itself the harness.

Sharing my agents.md folder structure with you all, broken into two images. I'm about two months into this, so I'm still learning. Curious if this looks bloated, if there's a cleaner way to set it up, or if anything here could be improved. Would love to hear what the community thinks.