AI Sprint: Claude Code with Antigravity — Day 1 (Taste as the Moat, the Muscle-and-Mirror Setup, and the Model Council Build)

The rumination he brings to class: Mythos-tier capability (Carlini's security findings; training-data operations where workers wear head cams so robots can copy them) makes 'sit still and hope' a guaranteed loss. His three survivals: TASTE — 'AI can give you a hundred options in ten seconds, but taste is where you figure out which — it's accumulated judgment'; SYSTEMS THINKING — 'figuring out what problem is worth solving in the first place'; CRAFTSMANSHIP — depth AI can't match. The sprint then practices what it preaches: no button-clicking instruction ('UI/UX, where the button is, is something you figure out yourselves'), maximum attention on judgment calls — which skills to trust, which repo signals matter, when to double-check the agent, what to commit.

The stance is offered with epistemic honesty — 'I don't know if I'm right; this is genuinely what I believe' — which is itself the taste lesson: conviction plus humility, held publicly.

The architecture: CLAUDE CODE, terminal-first ('the purest form... a learning curve, but that is where it has the best capability'), installed as the official Anthropic extension into Antigravity (or any VS Code descendant — 'everybody has the same VS Code underneath; the setup is absolutely the same'). ANTIGRAVITY runs in EDITOR MODE ONLY — its agent manager deliberately unused — for the mirror function: 'whatever Claude is building, if you want to SEE it, you need an editor' — including hidden dot-folders (.claude, .env) invisible to the desktop. And the failover function: 'even if Claude Code tokens get exhausted, Antigravity will continue building, and you come back to Claude Code when your tokens reset.'

The drill sequence: terminal basics FIRST via codingformarketers.com ('spend at least one hour on this site before you touch Claude Code — internalize it so when Claude runs a command, you know what it's doing'; cd is the one command you truly need; PowerShell on Windows). Then the three commands: /login (pro/max plan or console credits), /USAGE — 'the command nobody guesses... on the pro plan, assume 20-25 minutes of building and your tokens are exhausted' — and /model, with the standing rule: 'for thinking and hard planning, use Opus. Once your plans are ready, use other models. If you use Opus for everything, you will lose everything in 10-15 minutes.' The capstone axiom: 'the most efficient token usage is possible only if you have CLARITY... that is the biggest token saving overall.'

Then the live test: Claude errors out mid-build ('I think in Dubai, in US, a lot of places, Claude has gone down'), he opens Antigravity's agent manager — 'go through the files and catch up on what is happening' — reinstalls the repo-scoped skills there, continues, and swaps back when Claude returns. The md-and-files-on-disk architecture is WHY the handoff costs one sentence.

The frame: SKILLS are recipes — 'playbooks I use again and again'; MCPs are the TOOLS that cook them; the CHEF is Claude Code, 'bringing the recipe to life with the tools.' The knowledge-gap doctrine follows: the best product managers have written their recipes (Pavel Huryn's PM Skills repo, carrying Marty Cagan and Teresa Torres playbooks) — 'will you be stubborn and rediscover it on your own, or stand on the shoulders of giants? I'll take my grandma's recipe and add my own taste on top.' Judgment gates the import: repo health read from ~1,100 forks and ~10,000 stars ('1,100 people copied it and build on top; 10,000 use it').

Then the move he flags as 'a very difficult trick — watch carefully': DON'T install everything. Paste the repo link, describe the actual build, and ask WHICH skills apply — the agent's answer is the lesson: 'Honest take: most PM skills here are for PM strategy work, not for building an app. Skip the plugins you don't need; install just these.' Install scoped 'for me, in this repo only' so setups don't clash ('I have a different setup; I don't want these to clash'). Vocabulary landed on the way: PLUGINS are collections of skills ('a skill for SEO, copywriting, landing pages — together, a marketing plugin').

The build, end to end: the create-PRD skill interviews him (which models on the council; debate mechanics — round 1 independent answers, round 2 cross-critique, optional round 3 final statements; a separate synthesizer judge; web-app form factor; verdict table with agree/disagree, confidence, reasoning trace, exportable markdown). His answers: Sonnet 4.6, GPT-5, Gemini 3.1, Grok latest, Opus as judge, all via OpenRouter. The taste moment arrives immediately: the agent claims 'Gemini 3.1 doesn't exist yet' — 'whenever you see such things, always double-check' — he opens OpenRouter's catalog live, finds Gemini 3.1 Pro preview and Grok 4.2, and corrects the agent with evidence. Design arrives as a copied prompt from Super Design ('clean, minimalist'); the assumptions skill audits the idea; the Claude outage interrupts; the app ships anyway.

The first run is a gift: asked 'is Claude Code better, or Antigravity?', the council HALLUCINATES — two models argue Antigravity isn't a real product, two invent descriptions of it — 'a genuine factual disagreement about whether the thing exists,' because the models had no web access. The fix is one detail: OpenRouter's ':online' slug suffix bakes web search into every model. Then the ship ritual: push to GitHub ('don't commit the PM skills — they already live on GitHub'; .gitignore verified; .env excluded), README written, repo public. The accounting closes it: $0.26 of OpenRouter credit, 28% of a max-plan session — against Perplexity's $200/month tier.

The full ritual, enacted rather than lectured: keys live in .env ('environment variables store your secrets — passwords and API keys — separately; you should NOT upload this to GitHub'); the agent creates the FILE but the human places the KEY ('try to avoid keeping your key in the chat'); the screen share pauses during the paste — the rhetorical question making the audience derive the rule; .gitignore is READ before the push ('git-ignore is very, very important — what files are getting ignored? Good, .env files are ignored'); and the commit scope is curated ('don't commit the PM skills — they're already available on GitHub; I don't want to commit skills that live upstream'). The trust posture is stated exactly: 'it knows what it should not do, but better to instruct' — verify the agent's hygiene even when it's probably right.