Whether something clicks and sticks is another question. But I've compacted the ICM process into six steps for my brain.

I’ve been working with ICMs, and this explanation finally made it click for me: A repeatable process is a strong candidate for ICM context. The ICM should preserve: • When and why to use it. • Required inputs and examples. • Steps and decision rules. • Quality checks and approval boundaries. • What worked, what failed, and where to resume. The reusable skill carries out the steps; the ICM supplies the context for choosing and using it correctly. That’s helping me understand software in a system. The tools can change while the method, context and lessons remain available to whoever picks up the work next. I liked this explanation enough to ask that it go at the front of our ICMs. Apparently, I also need context for my context. 😂 What repeatable process would you preserve first?

RinDig/icm-architect It runs in pretty much any agent now: Codex, Gemini CLI, Hermes, Claude Code, Cursor. Or hand the folder to any model and tell it to use the skill. Workspaces now start from an AGENTS .md file, so they aren’t tied to one company’s tools. It would work with others before just not as reliably, should be better now. I also cleaned it up a lot. The whole method comes down to three building blocks now: a sequence of steps, a shelf of same-shaped files, and a small map that points to everything. Every step ends at a check, whether that’s you looking at it or a small model doing it, so a mistake gets caught before the next step builds on it. There are also some small scripts that check your links and draw your whole workspace as a graph you can click through (very basic graph, obsidian is better for viz ill make a nicer one later) github.com/RinDig/icm-architect Install steps for each agent are in the README. Try it on something you do every week and tell me where it breaks.

@Jake Van Clief I am building Onyx Commercial Roofing's system for maintaining our SOPs via github like you show in training. Here is what I have been able to come up with https://github.com/adib-gmap/OCR-SOP-structure-and-processes.git. Note: This is not the actual SOP's this is the Frame work built to Manage the SOP enviroment in a productive and scalable manner for our team. I have been writting sops for so long and managing them is hearding cats. Think of the SOP's as a individual soldier. I am trying to create guide for the O3 on how to manage a Company of soldier so your Company is dress right dress, not just in formation standing around, but also when conducting operations, the guide is how the Company moves as a whole. What rules the E8 is responsible for enforcing, if he wants to change them, how so? Same for E6's, E5's ect. I have been able to identify some gap, for each gap below I put in a placeholder rule. Is it right, what would you do instead? Changes and approvals 1. Major vs minor: MAJOR = someone has to do something differently; MINOR = more detail, no new obligation; PATCH = typos. Does a major change force retraining? 2. Who approves: the team owner approves MINOR; team owner + business owner approve MAJOR; never the author. Right people? 3. Safety / warranty SOPs (OSHA fall protection, manufacturer install specs): I added a qualified-reviewer sign-off. Needed, or overkill? Keeping it alive 4. Review cadence: safety/warranty every 6 months, ops 12, admin 24, plus an immediate review after an incident or repeated deviations. How do you set the S11 maintenance schedule: time, usage, or only when gaps show up? 5. Field deviations: the crew calls the PM first if it's safety or warranty, otherwise logs it within 24 hours. At the weekly meeting each one gets an outcome: one-off, coach the person, or change the SOP. Is that how you'd feed the S11 loop?

Last week I posted part one of Second Eyes (https://www.skool.com/cliefnotes/i-expected-a-second-ai-from-another-company-to-catch-more-it-caught-less?p=4179abfc. Quick recap... I planted 288 mistakes in 12 renovation cost estimates and had Claude, ChatGPT and Gemini review them in the browser, 180 reviews in all. A second AI from a different company didn't catch more than a second Claude. Two Claude reviews caught 84%, and Claude plus another company caught about 82%. A few of you asked the obvious next question. What about the tools people actually use? So I tested mine. WHY THIS ONE My everyday setup is Claude Code for the work and Perplexity as my second set of eyes. Claude Code is the version of Claude that can write and run code, so it can recompute every number instead of reading it. Perplexity picks the model for you and can search the web. If any setup was going to beat "ask Claude twice," I figured it would be this one. WHAT I DID Same 12 sheets, same 288 mistakes, same prompt and answer key as part one. Claude Code reviewed every sheet 5 times with all my personal settings switched off. Perplexity reviewed every sheet 5 times in incognito on its "Best" setting. That's 120 reviews, all on the same day. Before any data came in, I wrote down what would count as a win. Perplexity as the second look had to catch at least 5 points more than a second Claude Code review. WHAT CAME BACK Claude Code plus Perplexity caught 83.8% of the mistakes. Claude Code plus Claude Code caught 83.2%. Half a point. The 5-point gain is ruled out. On its own, Perplexity caught about 72% and Claude Code about 77%. And Claude Code caught about the same as the regular Claude chat from part one, 77% against 79%, even though it ran Python on every single sheet. The part I found most interesting... Perplexity and Claude miss different things. Perplexity was better at catching work done in the wrong order. It was much worse at catching numbers that contradict each other across the sheet, and claims that weren't true. But when I paired them, those differences mostly cancelled out. A second Claude Code review picked up about as many new mistakes as Perplexity did.