First, I have to say it: Nate's video genuinely blew my mind. The roast skill changed the way I make decisions, and the whole "stop letting the model agree with you" idea is gold. If you haven't run it yet, do — it's that good. As I adapted it to my own workflow, I noticed two small additions that pushed it even further for me. Funny enough, they're the exact same two things that leveled up my multi-LLM council a while back: a second pass, and more than one brain. Tweak 1 — a second round (cross-examination). Round 1 stays exactly like Nate's: the six personas attack independently and score the idea. Then I added a round 2 where the personas look at each other's takes and pressure-test them: - Expansionist → Contrarian: is that "fatal flaw" really fatal, or are we over-worrying it? - Contrarian → Expansionist: is that upside actually reachable with the resources we have, or is it wishful? - Deep-researcher → everyone: flag any claim with no evidence behind it (anything unverified loses weight). - Buyer → everyone: does the real customer even care about this, or is it just an internal debate? - First-principles → everyone: separate solid logic from hope, and keep anyone from quietly redefining the product. Each persona revises its score after the pressure, and the judge synthesizes only from round 2. Anything that doesn't hold up in the cross-examination doesn't get to drive the verdict. It turned the output from a clean yes/no into a much deeper, debated call. Tweak 2 — bringing in other LLMs. Nate's version runs on Claude alone — one model playing all six personas. I wanted more than one point of view at the table, so I added other LLMs to the panel: Codex (GPT) and Gemini, alongside Claude. Now Codex and Gemini run the adversarial panel and cross-examine each other in round 2, while Claude handles the live web research and acts as the judge. Three different brains genuinely disagreeing gives me sharper feedback than one model talking to itself.

The concept: A panel of AI models that debate my questions and hand me one sharp answer — running at $0, on subscriptions I already pay for. The idea I started from (Andrej Karpathy's): Different AI models — ChatGPT, Gemini, Claude — often give different answers to the same question. So instead of trusting just one, I liked Karpathy's "LLM Council" idea: ask several AIs the same question, let them read and critique each other's answers, and then have one model act as "chairman" to merge it all into a single best answer. It's like getting a second, third, and fourth opinion — then a senior expert who weighs them all and decides. How it works (3 steps): 1. I send my question to every model at once → each writes its own answer. 2. Each model reads the others' answers (anonymously) and critiques them. 3. A "chairman" model reads everything and writes one final, judged answer. What I built: My own version, running inside my AI assistant ("Brain"). Same 3-step idea — but instead of paying per use, I plugged in the AI tools I already pay for. The clever part (how anyone could do it): Karpathy's original talks to all the models through a paid service (OpenRouter) that bills you per use. I swapped that out for: - Claude — already running as my assistant; it also plays chairman. - GPT-5.5 — through the Codex command-line tool, which runs on the ChatGPT subscription I already pay for. - Gemini — through a free Google AI Studio key. A small script ties the three together. Why I made it free — the key thing I realized: Paying for ChatGPT / Gemini / Claude as apps is not the same as paying for their API (which charges per word). But the command-line agent tools (like Codex) run on the subscription I already have. So by wiring up those tools instead of calling paid APIs, my whole council runs at $0 extra — no new bills. I was already paying for these AIs; now they work together. Why I think it's worth it: - Better answers — multiple viewpoints plus cross-checking beat any single model, and catch blind spots. - $0 extra — it reuses subscriptions I already pay for. - Private & local — it runs on my own machine. - Smart chairman — Claude does the final synthesis and already knows all my context (projects, priorities). - On demand — I use it for big decisions, like planning a major project, not everyday stuff.

Here's mine. I have a Telegram bot running on my phone. Any time I see something worth saving — a video from Youtube, an idea on social media, a thought mid-walk — I just forward it there. Takes 2 seconds. How it works: 1. Create a bot via @BotFather on Telegram — takes 2 minutes, gives you an API token 2. Write a small poller script (Node or Python) that calls getUpdates on the Telegram API every few seconds 3. Script writes every incoming message to a local file (inbox.jsonl) 4. Claude Code reads that file every morning as part of the daily brief skill 5. Actionable items land in Obsidian — as tasks, project notes, or ideas tied to active projects — and the inbox resets. Clean slate every day. No cloud service. No third-party app. Just a script running in the background on your machine. The insight: an AI OS is only as good as what you feed it. Most people spend weeks building the automation layer but never solve the input problem — so the system runs on stale context. The bot fixed that. Now wherever I am — away from my desk, walking, traveling — I can capture thoughts, interesting finds, and ideas on the go. Nothing gets lost. Curious how others are solving this — how are you feeding live context into your AI OS?

Hi guys, This might be a totally noob question, so forgive me, but I'm wondering what the best file structure for Claude Code is? When do you use global instructions and when do you use project-based instructions? If I'm connecting into a specific project folder, will it still read the global claud MD information and the business context? is it better to save skills globally instead of repeating them per project? My thinking is that I want to have a global layer where all of my context around my business, what I'm working on, any references, skills and everything are just kept in one place, and I'm not repeating it in every single project. And then on the project folder, it's only things related to that project specifically. Where should I be pointing when I'm working on VS Code? Should I be pointing to the project folder? Will it still read the global instructions, or should I always point to the global folder? What are your best practices or the best ways of handling this guys? Thanks!!

The hardest part of building a good AI system isn't the prompts, it's getting everything out of your head and into the system. In this video I break down the grill-me skill, which relentlessly interviews you about a process and writes it all back to a knowledge doc so nothing gets lost. I show how I built this to checkpoint after every answer, why front-loading context gets your skills to 90% on the first try instead of grinding through 30 iterations, and how to start grilling yourself today.