Activity
Mon
Wed
Fri
Sun
Nov
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
What is this?
Less
More
Build Market Close
719
5.0
Free
Clief Notes
51.9k
5.0
Free
68 contributions to Clief Notes
2d •
My AI Is Brilliant. So Why Am I Still Supervising It?
I'm curious, when you are working with AI (Claude or choose your model) and ICM, what do you find yourself having to remind your AI to do (or NOT do) most often? I mostly use Claude to build and it seems to be doing much better at reading the instructions, staying within scope, and not making assumptions, but every now and then, something off-the-wall still happens. I'd love to hear what has shown up in your workflows. Surely I'm not the only one who has had these conversations at one time or another.
1 like • 1d
I use Codex and because I have that on the paid mied tier plan mostly ChatGpt for Chats. Sometimes I use Gemini and NotebookLM. I’m currently building my Knowledge MVP and I named part 3 literally this: ICM Series 03 — I Opened Codex, Then Told It Not to Build Anything 😄 Especially during building I tell it to stop, do not create, edit, move, or delete files until I approve the proposal. I find Codex rather trigger happy. Not in the way of doing it wrong, but in the way of seeing a step ahead and trying to do something I didn’t want yet because it does not know yet what I wanted. It had worked well so far and it does follow the instructions. In case you’re interested, or need help falling asleep 🤣, I wrote about it here: https://www.skool.com/cliefnotes/icm-series-03-i-opened-codex-then-told-it-not-to-build-anything?p=35ab0abe
0 likes • 3h
@Danish Saleem Yes, so far I have and Codex has been obedient .😄 Making sure Codex doesn't hop forward to step 15 when you're ideating and testing step 1 was actually one of the things I was trying to avoid. What I've been doing is making a pretty clear distinction between suggesting the next step and having permission to execute it. I let Codex propose what it thinks should happen next, but usually add something along the lines of: "This is a proposal only, not approval to proceed. Do not implement the next step or make changes until I explicitly approve it." That's actually how the first kickoff worked. I told Codex to propose Phase 1 only and not create, edit, move or delete anything until I approved it. And when the proposal came back, I actually made it smaller before giving the green light. The other part is keeping the agreed decisions and current project state in DECISIONS.md and BUILD_STATUS.md. That way, a suggested next step doesn't automatically become the new mission. It's not some magical technical lock, but clear boundaries, explicit approvals and persistent project state have worked pretty well for me so far. Well, I hope I didn't jinx it now. 😅 And there's another interesting experiment around exactly this later in the series.
4d •
ICM Series 03 — I Opened Codex, Then Told It Not to Build Anything 😄
⬅️ 𝗣𝗥𝗘𝗩. 𝗦𝗧𝗘𝗣: 𝟬𝟮 — 𝗙𝗿𝗼𝗺 𝗖𝗵𝗮𝘁 𝘁𝗼 𝗙𝗶𝗹𝗲𝘀│ 🏠 𝟬𝟬 𝗦𝗧𝗔𝗥𝗧 𝗛𝗘𝗥𝗘 │𝗡𝗘𝗫𝗧: 𝟬𝟰 — 𝗡𝗼 𝗖𝗵𝗮𝘁 𝗛𝗶𝘀𝘁𝗼𝗿𝘆 ➡ Quick recap. In Step 01 I figured out I wanted to build a little Knowledge System MVP. In Step 02 I turned my ideas into a rough plan and moved the important thinking out of chat and into 𝗣𝗥𝗢𝗝𝗘𝗖𝗧_𝗦𝗣𝗘𝗖.𝗺𝗱, 𝗕𝗨𝗜𝗟𝗗_𝗦𝗧𝗔𝗧𝗨𝗦.𝗺𝗱, 𝗗𝗘𝗖𝗜𝗦𝗜𝗢𝗡𝗦.𝗺𝗱 and 𝗖𝗢𝗗𝗘𝗫_𝗞𝗜𝗖𝗞𝗢𝗙𝗙_𝗣𝗥𝗢𝗠𝗣𝗧.𝗺𝗱. Surely now it was time to unleash Codex and tell it to build the thing? Well... not quite. 😄 𝗙𝗶𝗿𝘀𝘁 𝗜 𝗰𝗿𝗲𝗮𝘁𝗲𝗱 𝗮 𝘀𝗵𝗼𝗿𝘁 𝘀𝘁𝗮𝗿𝘁 𝗽𝗿𝗼𝗺𝗽𝘁 𝘁𝗼 𝗸𝗲𝗲𝗽 𝗖𝗼𝗱𝗲𝘅 𝗶𝗻 𝗰𝗵𝗲𝗰𝗸: Please read 𝗖𝗢𝗗𝗘𝗫_𝗞𝗜𝗖𝗞𝗢𝗙𝗙_𝗣𝗥𝗢𝗠𝗣𝗧.𝗺𝗱 in the root of this project and execute the instructions in it. Before making any changes read the referenced project files. Use the globally available icm-architect skill as instructed. Propose Phase 1 only. Do not create, edit, move, or delete files until I approve the proposal. That was it. No giant prompt explaining the whole project again. The files already held the context. The message mainly told Codex where to look, what to do and where to stop. 𝗘𝗻𝘁𝗲𝗿 𝘁𝗵𝗲 𝗜𝗖𝗠 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁 — 𝘁𝗵𝗮𝗻𝗸 𝘆𝗼𝘂 @Jake Van Clief 💪 I had already made the ICM Architect available as a reusable skill outside this project. But I didn’t want to just use it straight out of the box. I wanted to use it like a good consultant: listen first, then work within what the customer actually needs. --> Codex was the worker. The ICM Architect was the specialist it consulted. Codex first had to read the project status, decisions and relevant parts of the project spec. So the Architect worked inside the requirements and boundaries I had already established, not instead of them. I quite like that pattern: bring in specialist capability when you need it, but keep the project in charge.
2 likes • 1d
@Ron Davis Love this, I do the same things. I spent lots of time going back and forth, changing things here, taking in the feedback, adding stuff, subtracting stuff and then at some point it's a well summarised handover file. I suppose it's the 60/30/10 in action isn't it? 🙂
1 like • 20h
@Dino Behler thank you so much. I’ve done lots of consulting in manufacturing, although I actually dislike the word consulting. I always say I’m a problem discoverer and solver. 🤣 But that’s when I saw the parallels and thought wait a minute. Glad it resonates 🙂
1d •
ICM works too well!
I have 3 OAI resets 1 Anthropic reset Codex is at 1% of weekly total used Claude at 9% I still have Fable usage credits left 1000 on Codex. How am I supposed to use up all my resets if my folder structures are doing the job too well I have 4 active research projects lol
1 like • 1d
🤣 That's a great problem to have isn't it. I'm the same boat at the moment despite building ICMs. We have to consult @Ari Evergreen about this one. I heard she can use a few more tokens although our extra resets might only last 5 minutes with the stuff she's working on.🤣
3d •
Introduction
Hey guys, I'm Matt, an indie designer/developer in LA. I run a solo studio (client work and my own products) and a local security camera business. Besides those I'm usually tinkering with local LLMs, 3D printing, or robotics. I randomly came across one of Jake's youtube videos not too long ago and was about to close the tab expecting the usual guru-slop content. But when I saw his desktop wallpaper, I knew I had to watch all his videos. (If you're into DBZ we're already friends) So now I'm here and glad I joined! I already learned a lot but I'm trying to figure out how to organize my folders for 2 businesses and a bunch of side projects. Has anyone else here struggled with that? Would appreciate some help. Anyway, I look forward to learning, building, and connecting with you guys. Hit me up if you're a fellow designer, developer, DBZ fan, or if we share any interests like 3d printing, etc.
1 like • 2d
@Jordan Shaw The Hormozi advice is spot on 😄. It could save you a lot of work. Thanks for the detailed response. Just to clarify, I’m not actually dealing with multiple businesses myself. I’ve been building an ICM-based Knowledge System MVP, and a lot of what you’re describing is very relevant to my project. Some of your experiences reinforce decisions I’ve already made. A bit of confirmation that I’m at least not going down the totally wrong path. I’ve implemented provenance tracking and access controls, and have been working on keeping knowledge separate from the processes and models working with it, rather than relying purely on instructions. However, you’ve given me something new to think about 🧐. It’s one thing to keep knowledge separated, but another to allow controlled access across different areas while preserving ownership and provenance. Thanks for bringing that up. Definitely something I’ll incorporate as my system develops. I also appreciate your honesty about Hermes and its learning curve. I’m drawn to it, but it’s useful hearing from someone who’s actually been working through these problems. Helps with the FOMO 😜
0 likes • 1d
@Aurel Babiš Thank you for sharing this. That’s a very good test and actually quite close to some of the tests I’m planning with my “AI developer” for the next phase of my Knowledge System MVP. At the moment, the controls are primarily in the deterministic Python tooling and retrieval/evidence packaging. The AI can initiate a search, but the actual retrieval happens through a SQLite catalogue, with selected evidence packaged and linked back to its original sources. I’ve also implemented provenance tracking and some policy-based controls. My background is in manufacturing, and I’ve essentially borrowed the concept of a manufacturing traceability system. It works in a very similar way, just with different terminology. For example, manufacturing genealogy becomes provenance, and sidecar metadata serves a similar purpose to some of the information you'd find in a batch traveller. So I'm also doing translation work. 🤣 What I haven’t proven yet is stronger isolation at the filesystem/execution level. That's what I'm working on next, including testing whether an agent can actually list or read protected files rather than just being told not to. I really like your per-task grant idea. I already have task and evidence identifiers, but linking an explicit authorization grant and an audit trail of actual file reads to those would add another useful layer. That's something I'll take into account as the system develops. Thanks for the tip. 💪 Interesting to hear you're working on this with VaultGuard. Is the per-task grant mechanism already enforced for external agent sessions, or is that the part still being developed?
1 like • 1d
🤣🤣🤣
1-10 of 68
@stephan-vollmer-4733
Transforming manufacturing in all sectors for 18+ years. Love to build, experiment and learn new things.
Active 4m ago
Joined Sep 4, 2026
Powered by