Is there anyone else in this group using AI to build an app? If so, got any tips on folder structures and markdown files? I don’t know if what I’m currently doing is good ICM practice or if ICM is a good strategy for building with a code agent rather than doing things with multiple agents, as others in this group are doing.

@Ry Mac did a GREAT job simplifying things down. Recommend taking a look. Has his own group too! https://skool.com/buildmarketclose/about

I just saw Jake’s instagram post about this and i need some clarity. I’m about to start using Xcode with Claude and it mentioned “Claude Agent SDK” when I set it up. I used my Anthropic credentials to set it up rather than my API key. Is this setup going to require the separate usage from my Pro plan as well? If so, what is a setup I should be using to still use my Pro plan usage instead, like it does when I use the Claude Code CLI in my VS Code terminal?

💰 $325 CASH 💰 That's a full year of Premium. Win this and your membership pays for itself. But the real prize this week isn't the cash. Keep reading. 📋 THE CHALLENGE You just got hired again. Meet Diana, owner of a 4-person boutique real estate team in Austin. 60-80 transactions a year, mostly residential, mix of buyers and sellers. 📎 Download the full client brief attached to this post. Short version: She doesn't want software. She wants a system she can teach her team to use in a week. Your job is to build the AI operating system for her team. This isn't one specialist. This is a small team of AI specialists organized into a multi-folder ICM architecture, with a clear handoff protocol between them. 🗂️ WHAT YOU'RE BUILDING Last week was one specialist. This week is a team of them. Required folders: 📍 00_orchestrator/ — The router. Where every request starts. Decides which specialist gets the job. 📍 01_lead_qualifier/ — First contact with new prospects. Captures intent, budget, timeline. 📍 02_property_research/ — Deep research on specific properties or neighborhoods. 📍 03_client_communication/ — Drafts emails, texts, follow-ups in the voice of the agent. 📍 04_transaction_coordinator/ — Handles the deal once it's live. Checklists, deadlines, document tracking. Each folder must include: - 📄 identity.md - 📐 rules.md - 💬 examples.md - 🔗 handoff.md (NEW for Week 4 — how does this folder pass work to another folder?) - Plus a root-level README.md explaining the architecture, the typical flow, and how to onboard a new team member. 🔥 WHY THIS ONE IS DIFFERENT Weeks 1, 2, and 3 were warmups. This is the comp where the work you ship genuinely starts to look like the real thing. The handoff protocol is the test. Anyone can build five folders. The hard part is defining what each one needs from the previous one and what it passes to the next one. That's where multi-agent systems actually live or die.

New benchmark out of Meta FAIR, Stanford, and Harvard called ProgramBench. The setup: you get a compiled executable plus its docs. Source code stripped. Rebuild the program from scratch in any language you want. Tests check input/output behavior against the original binary. 200 tasks, from small CLI tools up to FFmpeg, SQLite, and the PHP interpreter. 📊 Results across 9 models: Zero tasks fully solved. Opus 4.7 was the best, passing 95% of tests on only 3% of tasks. GPT 5.4, Gemini 3.1 Pro, and Haiku 4.5 hit 0% in that bucket. The interesting part is section 5. Even the model solutions that "worked" looked nothing like the human reference. Median 1,173 lines vs 3,068 in the original. Flat directories. Fewer functions, each one longer. GPT 5.4 wrote 96% of its final code in a single turn on most tasks and never modified existing files on roughly 40% of runs. 🎯 Why it matters for us: The benchmark separates writing code from designing software. Models can produce syntax all day. They cannot yet decompose a real system into coherent modules, pick the right abstractions, or organize a codebase the way a working engineer would. That gap is what computational orchestration points at. It is also where the durable value lives. 🛠 Try it: Pick an easier task from the repo (the paper flags nnn, fzf, gron, and jq as more tractable). Run it against Claude or your model of choice. Watch where you and the model split. Note the design decisions you make that the model never even raises. Post your runs and attempts to create a harness that would allow the model to do it. Wins, failures, weird outputs, all of it. 📍 Paper and Repo: ProgramBench I'm building something on top of this right now. More soon.