Hi @Jen Cortez-Walters, Step 1: Take a deep breath. Step 2: Walk away from your computer and go do ANYTHING else for as long as it takes until your mind feels clear and the agitation/frustration fades to just a background noise. Step 3: Read my unnecessarily verbose response below. Do you have memory turned off as a setting? how are you working in sessions? Are you comfortable using VS Code? Would you mind sharing your CLAUDE.md? I'd even ask to share the workspace you are having a problem with. Honestly, my first step is a fresh session in the workspace and my first prompt is: Tell me everything loaded to session context with this init. Provide file paths where applicable and otherwise provide navigation/direction of the sources. In my first 2 months in Claude Code, this is where I was finding the most issues. I unknowingly let Claude load too much unrelated junk into a session and it was contaminating every with unnecessary context. If any of your routing in your workspace is "too ambiguous" or easily misinterpreted by the AI, it will overload or confuse its own context. As an example, for my comp 9 submission, I referenced workspace paths where rules were defined for certain elements. This led to those sessions pulling those docs into context when they weren't necessary and not part of the input files. I removed the path and modified the instructions in the stage so it didn't NEED the definitions because it just needed to handle the output of the prior stage. This was a mental somersault for me because I personally want to know "all the things, all the time" and it bled into the workspace. Another suggestion, start a session completely unmounted from your main workspace and make sure it can use the icm-architect skill Jake shared in this week. Invoke the skill and point it at the workspace you are struggling with. It might NOT capture every finding in a single pass, but it will capture the first things it runs into that break in your workspace. Iterate and fix with the architect skill across multiple sessions. Remember to test routing in the workspace as you work through the fixes so things are applying as you intend. Base on your share, it reads like you may have some systemic routing/direction issues.

I prefer second person instructions. The AI is "you" because when I am in a session that is also how I refer to it. Whenever Claude/ChatGPT/Gemini read instructions it appears to take "you" as addressing it, the AI.

When you don't read the instructions to move the folders despite staring at the arrows while you think... ICMadness · day 4 watchmaker-case-study · 3 stages 🟥🟩 solved in 2 runs · 20 moves · 5:11 donsbookshelf.com/icmadness/

@Don Roy Also, anyway to allow people to play prior days?

Repo: https://github.com/wcvessels/fiber-copy-editor Pure ICM play this week. Check out the README.md! Fiber Copy Editor reviews the business-to-business (B2B) fiber/telco marketing copy my team produces: blog posts, product pages, landing pages, social. I built it for my own marketing department and use it for real, a thinking partner that diagnoses but never rewrites. It points at the exact lines that fail, names the audience and purpose they fail for, explains why, and hands the draft back for the writer to solve. The specificity is in the granularity: it calibrates to each piece's exact audience, funnel stage, and format rather than giving one-size feedback. Eight passes run every draft through the full gauntlet: audience-fit, claims, proof, differentiation, structure, and call to action (CTA). Each problem is tagged a strategic higher-order concern (HOC) or a sentence-level lower-order concern (LOC), then the round is triaged into an edit letter and line-anchored notes. Its diagnosis is built on the messaging frameworks my team works from, Stratton's Make It Punchy and Dunford's Obviously Awesome, and I live-tested it with them. The voice/tone toggle was their request, a fun one to add. Next up: markdown intake so large PDF or Word drafts stop eating the context window.

@Adam James Technically, you could replace context.md with a more complete Claude.md at root. The issue becomes that other AI don't generally read Claude.md to start the way Claude does. So if you migrate your ICM, that system likely will not read the Claude.md to automatically get context for the workspace.