⬅️ 𝗣𝗥𝗘𝗩.: 𝟬𝟱 — 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝗜𝗻𝗴𝗲𝘀𝘁𝗶𝗼𝗻　│🏠 𝟬𝟬 𝗦𝗧𝗔𝗥𝗧 𝗛𝗘𝗥𝗘│　𝗡𝗘𝗫𝗧: 𝟬𝟳 — 𝗖𝗼𝗺𝗶𝗻𝗴 𝗦𝗼𝗼𝗻 ➡️ Note: Press on the links above to navigate to each series or home. At the end of Post 05, my Knowledge System was able to process documents, extract their contents and recognise unchanged files. A great achievement and another step forward. However, there was a bit of a problem. Just because you have the information you need doesn't mean you can find it. Imagine thousands of documents. Somewhere in there is the information you are looking for. But which file does that information sit in and where exactly within that file? It's the needle in the haystack. I certainly didn't want to ask an LLM to read everything each time. Remember, I don’t have tokens burning a hole in pocket. 😅 🔎 𝗦𝗼, 𝗵𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗳𝗶𝗻𝗱 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝘀𝘁𝘂𝗳𝗳? From Post 05, I already had the extracted content stored in JSON files on disk. Those files contained the information extracted from the original Word, Excel and PDF documents, along with details of where it came from. Technically, those files could be searched directly. However, with potentially thousands of documents in mind, I wanted a better way to organise and search that information. That's where SQLite came in. Essentially a lightweight database that could act as a searchable catalogue. The idea was simple: 📁 Original documents (Word, Excel, PDF) —> ⚙️ Extracted JSON files (stored on disk) —> 🗂️ SQLite catalogue (searchable index) —> 🔎 Matching passages and their source locations. Instead of going through all the extracted files for every search, SQLite could help locate matching information through its search index. Most importantly though, this was still deterministic processing, not AI reasoning.

Time for a quick intro! Hey Clief Notes fam! 👋 I’m Aaron, and I’m stoked to be here. 🙋 **A little about me:** I work in marketing, operations and automation. I enjoy figuring out how to make things simpler and more efficient. 🎯 **My current goal:** Get more done with less back-and-forth, build reliable systems, and free up more time for my family. 💪 **What I’m currently building/working on:** Putting AI to practical use in my day-to-day workflow. My focus is getting simple, useful solutions finished without turning every task into a massive project. 🤔 **My biggest struggle or question right now:** Knowing when to stop planning and ship. How are you getting consistent results from AI without spending all your time managing it or adding features you don’t actually need? Let’s get it! 🚀

𝗡𝗘𝗫𝗧 𝗦𝗧𝗘𝗣: 𝟬𝟭 — 𝗛𝗼𝘄 𝗜 𝗦𝘁𝗮𝗿𝘁𝗲𝗱 ➡️ (click it, it's a link 🙂) I’ve become more and more aware recently that there are quite a few people interested in ICM who get stuck on one fairly basic question: Where do I actually start? And I can understand why. There is a ton of really good information in this community, but it’s also a big community now. Add the way Skool shows posts, comments and search results, and sometimes finding the right thing at the right time isn’t that easy. So I thought I’d put together a little series showing how I started with ICM and how my own setup gradually developed from there. This isn’t meant to be “the way to do ICM.” It’s just another perspective alongside all the good material that is already here. What I looked at, what I tried, what worked, what didn’t, where I made mistakes, and why I added things as I went along. My example is a small Knowledge System MVP. It started quite simply and gradually grew into something with file ingestion and indexing, traceability back to sources, deterministic Python tooling, human review and control, fresh AI sessions being able to pick up old work, and a few other rabbit holes along the way. 😄 Some of the later stuff gets a bit more technical, but that’s actually part of what I want to show: you don’t need all of that to start. A lot of it only appeared because my particular use case eventually needed it. There’s also been a slightly funny Russian-doll effect along the way — using ICM to help build an ICM system, and then using that system to test some of the same ideas it was being built around. Because Skool unfortunately doesn’t give me a nice folder structure for all of this, consider this post my README.md. 😄 I’ll keep updating it with links as I add new posts, so if something disappears down the feed you can always come back here. Series 👉 01 — How I started

I see this question a lot, so I wanted to share my take. ICM (the stage-by-stage folder setup, where each stage loads only the context it needs) cuts waste. It doesn't cap usage on its own. If your usage is still high, it's usually one of these: **1. Your context files are too big.** This is a symptom of ai generated slop without proper human auditing, maintain your files! If each stage loads a 5,000-word reference doc "just in case," you've moved the bloat into a folder. Check what each stage actually reads. Keep stage instructions short and push reference material into files that are only loaded when needed. **2. Stages aren't isolated.** ICM only saves tokens if each stage starts clean. If you run everything in one long conversation, the model re-reads the whole history on every turn. Start a fresh session per stage and pass work forward through output files, not chat history. **3. Too much gets loaded at the start.** A bloated root `CLAUDE.md`, or a pile of always-on rules, skills and MCP tools, is paid for on every single message. Trim the always-on layer to the essentials. Manage your hooks and permissions properly! **4. Outputs are being re-read in full.** If stage 3 reads the entire raw output of stages 1 and 2, you're paying for it again. Have each stage write a short summary or handoff file, and have the next stage read that instead. My /handoff skill is one of my favorite and most used. One of the few things that's persisted in all my "restarts". **5. Your model choice is off.** Running the strongest model on every stage burns usage fast. Use a lighter model for mechanical stages (formatting, extraction, file moves) and keep the heavy model for the stages that need judgment. !! I was just burned on this today. I had my fancy new process all shiny and polished. Yesterday I was polishing and left a mark - today, Opus and Fable CRUSHED my account usage before I could notice, and I feel like I'm a careful person. **6. Loops and retries.**

Serious question for everyone running Claude, ChatGPT, AI agents, or automated workflows: **What if I could help you reduce your token consumption by 30–70% without sacrificing output quality or productivity?** Not by simply using fewer features, but by analyzing and optimizing your actual system architecture, agent orchestration, context management, model routing, and workflows. I'm developing something called **THE GERMAN STANDARD — German Agentic Systems Assurance (GASA)**. A structured technical audit and optimization framework based on established engineering principles and aligned with relevant international ISO standards. The idea is simple: 1. Audit your existing AI setup. 2. Identify inefficiencies, bottlenecks, and unnecessary token consumption. 3. Establish a measurable performance baseline. 4. Optimize your system. 5. Run identical benchmark tasks before and after. 6. Document the actual savings and verify that quality hans't declined. The 30–70% range is a proposed optimization target, not a guaranteed result. Actual savings would depend on your setup and would need to be demonstrated. **Now here's my question:** If I could demonstrate a measurable 30–70% reduction in token consumption on your actual workloads, with comparable results, what would that be worth to you? Would you prefer a one-time audit fee, a complete optimization package, or a performance-based model where I receive a percentage of the money you actually save? I'm not asking what you'd pay for another AI course. **I'm asking what you'd pay for a measurable improvement to a system you're already spending money on.** Would love to hear actual numbers.