I must say the new antigravity-cli (the successor to gemini-cli) is BLAZING fast and extremely good on Gemini 3.5 flash. And of course it can re-use your existing skills/tools from other agents. Doesn't mean I'll start using it, but really worth trying out.

Tim asked me the other day about Claude getting YT video content. I haven’t had any problems with Claude Code or Jarvis doing this for me, but it relies on me searching YT. So, I just built a YT Info Diet. It runs locally using Apify (free plan) and my Claude Code subscription. It runs every 12 hours, excludes Shorts and send the summary with a reco to my Slack. I gave it a list of channels to “watch” YouTube Channel Intelligence Monitor — Setup Guide Auto-summarize new YouTube videos from your favorite channels every 12 hours, with watch/skim/skip verdicts delivered to Slack. No API costs (uses Claude Code subscription). What You'll Get Every 12 hours, the monitor checks your favorite YouTube channels. When a new long-form video drops, it: 1. Fetches the transcript 2. Sends it to Claude for analysis 3. Posts a structured intelligence brief to Slack with WATCH / SKIM / SKIP verdict, summary, key insights, and an actionable playbook 4. Saves a markdown report locally YouTube Shorts are automatically filtered out. --- Prerequisites - macOS (this guide; Linux works too with cron instead of launchd) - Python 3.9+ - https://claude.com/claude-code installed and logged in (claude command working) - https://github.com/yt-dlp/yt-dlp — pip install yt-dlp - An https://apify.com account (free tier is fine — used as transcript fallback) - A Slack workspace where you can create webhooks --- Step 1 — Install Python Dependencies pip3 install feedparser youtube-transcript-api slack-sdk pyyaml Step 2 — Get Your Apify Token 1. Sign up at https://apify.com (free) 2. Go to Settings → Integrations → API tokens 3. Copy your token (starts with apify_api_...) Step 3 — Create a Slack Incoming Webhook 1. Go to https://api.slack.com/apps → Create New App → From scratch 2. Name it "YT Monitor", pick your workspace 3. In sidebar: Incoming Webhooks → toggle ON

Token management is real. I've seen people burn $10 in 2 minutes without realizing it. Here's the framework: → NEVER use your expensive model (Opus) for heartbeat/monitoring. Use Gemini Flash or a cheap local model. That alone can save you hundreds per month. → Use sub-agents for simple tasks. Your main brain should delegate, not do everything itself. → Watch your context size. The bigger the context, the more tokens every single message costs. Keep it lean. → Set up token tracking. Run /status regularly. Know what you're spending. → Don't paste your entire chat history into every prompt. Be surgical with context. Realistic budget: If you're smart about routing, you can run a solid AI agent setup for $100-200/month on API. If you're sloppy, you'll hit $500+ easy. The golden rule: if it makes money, there is no budget. Do more of it. But TRACK what's making money vs what's just cool. Don't trip over dollars to pick up pennies.