@Arjen Stet

Kids don't try this at home, these three jokers are paying members @Allan Durhuus @Arjen Stet & @Liani Smith ... They aren't punished for sillyness when using Clief Notes AI

@Aaron Kruger you're going to have to start charging for this information. Solid advice

@Adam James that's good news we have a lot of ballers in the community LOL

@Andrew Kloc ICYMI Jake did a video on this ...https://youtube.com/shorts/VNY7vWAzYt0?is=jt6B5y15NFIQp4QQ

You don’t need an agent. You need folders. @Ali ELBaitam , you’re not missing a secret framework. The “that” Jake keeps pointing at is usually a well-organized set of files. Context, direction, gated responses, a human in the loop. Close to 90% of the time, that’s the whole job. Scripts are rare. Multi-agent stacks are rarer. Stop hunting for the agent. Start mapping the work. What is simpler than a folder? A model reading the right files at the right moment, writing something you can inspect, then waiting. That’s ICM. Not vibe-coded slop. Not another black box. Interpretable, transparent, auditable. The filesystem is the architecture. Johan’s right. Not every use case needs an agent. Your follow-up is also right. Name the problem first. Keep the scripts, workflows, and BPMS that already work. Put a model only where judgment actually lives. Two gaps get lumped together as “ICM isn’t enough.” They aren’t the same thing. • Orchestration problems — parallelism, retries, branching, multi-agent coordination, scheduling. Put Hermes, Buzz, or similar on top of ICM. The folders stay the context and the handoff layer. The harness moves work. It doesn’t replace the map. • ICM-specific insufficiency — the problem itself won’t sit still as a stable, sequential, file-mediated, human-inspectable stage decomposition. Real-time multi-agent loops. High-concurrency serving. Automated mid-pipeline branching the system has to decide. In those regimes, filesystem-as-context is the wrong abstraction — no matter how fancy the orchestrator is. ICM plus a capable orchestrator covers a very large fraction of sequential and semi-parallel knowledge work. The residual cases above are the ones that still genuinely need something else. Don’t force folders onto that. Don’t duct-tape an “agent” onto a job that already has a deterministic pipeline. Garage-tinkering disclaimer: this is a rule of thumb, not production gospel. If your job is already exact, repeatable, and already coded, leave it coded. Leonard’s “when not to use AI” post in here is the same instinct.

@Ali ELBaitam In that case, definitely check out https://www.skool.com/cliefnotes/classroom/c7f102c7?md=9a0d198e1c0140578f37f1c93873c2b6 which clearly defines the terms that are thrown about. Claude = harness. You'll also find many beginner friendly resources to help you on your journey. Good luck

@Alex Brown Paper is a solid conductor-seat move. I just don’t treat the wait as dead time. When a job’s running I switch workspaces and start the next feature. Herdr is tmux for agents — when the running task needs me, I get pinged. Git worktrees keep the checkouts isolated, so two builds don’t collide. Watching teaches you nothing. Interrupting usually costs a rerun. A second workspace is the fourth option. Haven’t wired a tablet into a folder. That’s still the unbuilt piece — notes that never hit the workspace never hit the next prompt.