Hey Academy, Anthropic just made a major change to Claude Code—you can no longer use the Max subscription with the -p parameter or the Agent SDK. If you've built automations or your own AI agents on top of Claude Code, this affects you directly. In this video I break down exactly what changed, why Anthropic made the move, and what it means for developers relying on the SDK for automated workflows. I'll also walk through a clever workaround called Claude Heartbeat that uses Claude Code hooks to inject messages into your interactive session—essentially keeping your automations alive. Along the way, you'll learn how Claude Code hooks actually work and see practical ideas for using them in your own projects, like auto-committing code to GitHub when a session ends.

I grew an account from 6 to 23,000 followers in the past 6 weeks. Not by following the same recycled advice everyone posts. Most people are still running 2023 playbooks and wondering why nothing lands. Here's what I've found: 1. Pick one platform and optimize for it. The platforms are wildly different in what performs. I've seen creators pulling 100k average views per video on Instagram get literally 1,000 views on the same exact video on YouTube Shorts and TikTok. And vice versa. The content that wins on each platform is not the same. You should still cross-post because it's basically no extra effort and you might get traction depending on your niche. But unless you have a multi-million dollar production team cranking out platform-specific content like the Hormozis and Gary Vees of the world, pick one platform and go all in on it. The creators winning right now aren't "everywhere." They're dominant on one platform and letting the others collect scraps. 2. Skit formats are massively underused. Depending on your niche, skits are one of the highest-performing formats right now. We made a video for a franchise client comparing Wingstop vs. Chick-fil-A, just a simple skit breaking down the upsides and downsides of each. Super simple editing, super simple setup. It hit 5 million views. You don't need a production crew. You need a format that people actually want to watch. 3. Match your vibe to the platform. This is where most people screw up. On Instagram, higher-end, more produced content tends to win. Better graphics, polished captions, tighter editing. On TikTok, that same content usually tanks. What works on TikTok is raw. It should feel like a FaceTime call or a video message you'd send a friend. Basic text hook, basic TikTok captions, no fancy production. YouTube Shorts leans more toward produced content too, but it skews more mass-market consumer. Straight business content doesn't hit as hard there. 4. Use ai for certain parts of your content process (but the right way)

It's not just answering questions — it's showing you. "Hey, show me the best coffee shops near the Eiffel Tower" And it just... does it. Pans the map. Zooms in. Highlights spots. Talks back to you about what it found. This isn't search. This is conversation. --- What is this thing? Voice AI Conversation with Google Maps — a voice-powered AI agent that actually controls Google Maps while you talk to it. Ask it anything: - "What's the fastest route avoiding highways?" - "Find me beachfront hotels under $200 in Portugal" - "Show me where all the national parks are in Utah" - "Zoom out and show me the whole country" It doesn't just answer. It shows you. In real-time. While talking back. --- Why this hits different We've all typed into Google Maps. But talking to a map that responds, moves, and explores WITH you? That's a completely different experience. It's like having a knowledgeable local guide who also happens to control a giant interactive map on your wall. The Open Source Repo: https://github.com/moshehbenavraham/chat_with_google_maps --- The nerdy part (for my fellow builders) 🔧 Started with Google's AI Studio demo — cool proof of concept, but basically a toy. Nearly FULLY AUTOMATICALLY built: - Full authentication system - Secure backend (no exposed API keys 🔐) - Database for saving your stuff - AI monitoring dashboard and logging - Local/Dev/Production deployment on Vercel - 18,500+ lines of code Built the entire thing using the Apex Spec System — an open-source spec-driven Plugin/Skill for Claude Code. Every feature was spec'd out first, then built systematically. Complex project, zero chaos. 🔗 github.com/moshehbenavraham/apex-spec-system --- The future is conversational We're moving from: Click → Type → Talk This is just the beginning 🚀 Voice + AI + Maps is one combination. What about Voice + AI + your industry? The patterns we built here — security, monitoring, auth, database — they're reusable building blocks.

I run a UK digital agency and one of my clients is a builder who sends me raw build project videos via WhatsApp. We turn these into branded shorts and reels for YouTube, Instagram, Facebook, and Nextdoor. Current setup ($100/month): My VA in Nigeria handles the whole pipeline manually: • Downloads videos from WhatsApp • Runs them through Descript (adds voiceover when client’s ops manager hasn’t recorded one) • Polishes clips in Opus Clip Pro • Adds logo and branding • Writes captions • Schedules to YouTube, Insta, Facebook • Manually posts to Nextdoor via GHL workspace He’s reliable, the quality is OK, and at $100/month it’s hard to beat on cost. The AI-automated alternative I scoped: Make.com + Claude API + Creatomate (branding) + ElevenLabs (voice clone) + Opus Clip Pro API + GHL scheduling. Roughly £60-75/month in tooling, plus probably 20-30 hours of build time on my end. The build would handle: WhatsApp ingestion, Claude analysing each video and writing platform-specific captions, voiceover generation when needed, automated clipping, branding overlay, scheduled posting across all platforms, with one approval gate before publishing. Where I’m stuck: On paper the AI route looks cheaper long-term, but when I actually compare: 1. My VA costs $100/month flat, no build time, no API surprises, deals with edge cases naturally 2. The AI stack is ~£60/month in tools but needs my time to build and maintain, and breaks when APIs change or videos are unusual 3. Nextdoor still needs a human click either way (no API) 4. Voice cloning needs careful setup and consent 5. The VA can also handle other ad-hoc tasks the AI can’t My questions for the group: • Has anyone built something similar and found the maintenance cost was higher than expected? • Is there a hybrid worth considering — AI for the captions and Claude analysis, VA for the video editing and posting? • Am I overcomplicating this when a good VA with the right SOPs is genuinely the more cost-effective answer?

Quick version in case you missed it: instead of using RAG to re-chunk your sources every time you ask a question, you compile each source once into a persistent markdown wiki. The LLM extracts concepts, writes entity and concept pages, updates cross-references, flags contradictions, and maintains the whole thing. Future queries read the pre-synthesized wiki. The part that clicked for me: the reason most of us abandon our second brains is that backlink and cross-reference upkeep is boring. The LLM doesn't care. It's happy to touch fifteen pages in one pass. I spent a couple of weeks turning Karpathy's pattern into a Claude Code plugin that actually scales (atomic pages, sharded indexes, BM25 fallback past ~300 pages). It also runs in Codex, Cursor, Gemini CLI, Pi, and OpenClaw through the skills CLI. Install in Claude Code: /plugin marketplace add praneybehl/llm-wiki-plugin /plugin install llm-wiki@llm-wiki Or in any other supported agent: npx skills add praneybehl/llm-wiki-plugin -a <your-agent> Five slash commands (init, ingest, query, lint, stats), stdlib-only Python, no dependencies. Plays well with Obsidian if you want the graph view. Repo: https://github.com/praneybehl/llm-wiki-plugin Karpathy's gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f Curious if anyone here has tried the pattern themselves. What did you ingest first, and what broke before it worked?