Something I’ve been trying to wrap my head around is Jake’s Edubaware. Like most learning exercises, I start with a Google NotebookLM of the top sources on the topic. This time I had it build a slide deck overviewing Edubaware, then exported the whole thing as a PDF. Attached. The short version: it’s multiplayer folders. You take an ICM (a folder of files that already tells the model exactly what to do), freeze it as a template, and launch a workbench. Everyone on the team works against that same shared copy. The agent knows who is typing. No more hunting through separate chats for the good version. It’s basically the same shape as opening a project folder and running a few AI sessions against it—except the folder is shared, the template stays frozen unless a builder updates it, and you control the org, the roles, and the keys. Upload the zip that already works on your machine. It works for everyone you invite. Thought I should share the deck. Let me know what jumps out.

Same wall, every time. I ask a model for docs and I get one blob that serves nobody. Too fuzzy for an agent to follow without drifting. Too stiff for a person to read without squinting. And when I ask for a diagram or a video on top, the wrong claims just get prettier. So I built doc-clarity. One topic in. Writing you can check. Then a climb, one rung at a time, as far as the brief asks: text → diagram → web page → explainer video. https://github.com/dvogeldev/doc-clarity Rung 1: the writing. Three separate files, each under a real standard, because an agent and a person should never share one mixed document. 1. machine/text.md, full ASD-STE100. Controlled vocabulary, one meaning per word, short sentences, active voice. This is the file your agent or your code reads. It's deterministic, no "did the model mean this or that", and a checker script (ste_check.py) audits it. It writes pass or fail. Fail means the agent fixes the text and runs it again. 2. ste-human/text.md, STE at 80%, only when you ask. Most of the discipline, but a person can actually read it. Technical precision with humans in the loop. 3. human/text.md, ISO 24495-1 plain language. Four principles: relevant, findable, understandable, usable. The test: after one read, the reader can do the thing. Rung 2: the diagram. A hard paragraph becomes a figure. One claim per figure, and no new claims. The figure shows exactly what the writing already says, nothing more. You get an editable source file, not a baked PNG you can't fix. Rung 3: the web page. One self-contained HTML file. Interaction only when it helps the reader do the task. It's discardable, cheap enough to make for one topic and throw away. Rung 4: the explainer video. The narration and the shot list get written first, and they're checked against the writing before anything renders. Claims match, order matches, then render. No API keys stored. Why the order is fixed: 📐 A wrong claim is cheap to fix in text and expensive inside a video. Every later rung repeats the checked claims. It never extends them.

I think I had a fundamental misunderstanding of ICM. I was going through the course to go through the course, trying to build an ICM Builder without a clear product to build. Working through the old competitions is changing that. Starting with #4, The Agency, we built a five-role mock around Diana’s real-estate brief. Now there’s a purpose and something concrete to check. In one fictional case, I approved a budget change from $500,000 to $450,000. A fresh role resumed from the saved information, preserved the June timeline, and asked which year June meant. I reviewed the draft, but nothing was sent. Tests also held the review gate when approval was missing and when source text tried to override instructions. This is practice, not a finished real-estate system. All five roles have task-level checks; full-team onboarding and broader reliability remain open. The education for me is learning to build for a purpose before deciding how many systems I need. Later, I can apply this to my YouTube or studio ideas and ask what’s actually useful, and what’s redundant. I’m doing them for the learning as a prize.

Most of us run AI like a vending machine. Prompt goes in, product is supposed to come out. Instead the output drifts, the context window fills, and everything collapses. So we blame the model and try again. The article's punch: the model was never the answer. The structure around it is. Systems engineering kept a nuclear plant stable on a submarine, and it is the same loop that finishes pool jobs on budget, survives nonprofit handoffs, and carried me 42 days on a bicycle from California to Michigan. The loop, everywhere: 🔹 Research. Know the outcome before you touch the tool. 🔹 Declare. One page, in terms you can measure. 🔹 Cycle. Run small. Check. Track. Adjust. Add nothing yet. 🔹 Layer. Only after the core holds. Karpathy framed it for agents: fill the context window with exactly the right information for the next step. Your move: pick one system this week, write its one page, run the first cycle, and drop the page below. The comments are the course. The text is fuel. https://www.skool.com/cliefnotes/classroom/c7f102c7?md=2b72dd8a1c464a238cf995646d8b40bc #systemsengineering #contextengineering #auditable #ownYourStack #CliefNotes

Now that I have around ten ICM projects active and running everyday, it’s getting harder and harder to manage. Some accidentally were overlapping with other routine functions for example. Most, if not all of these are triggered by routines in Claude. With this many running it’s getting harder to manage all of it. I can get some real examples but I’m getting this feeling of being overwhelmed a lot lately, there’s so many things running and the more workflows you have running, the more issues pop up, problems to solve, adjustments, edits, etc. They all run from one master folder but more and more they seem to not be connected or working from the same playbook. It’s hard to explain. Most of my day now is reading long, dense blocks of text describing the results of that days run and issues from each routine, then spotting issues and correcting them, then moving to the next. It’s like spinning plates, one slows down so u go to spin that one and when it’s fixed, another one starts slowing down and the cycle repeats. I don’t have time to build anything new because it’s so much maintenance. Anyone else having this same issue? What’s the fix? I can provide examples or more context if needed. Any advice would be great:)