The day you build something.
The Big Three. What each is actually for. What to pay for, what to get free.
An LLM with a goal, in a loop, calling tools. Why that one idea broke everything open.
A real Chrome extension. Yours. Installed on your machine by the end of the day.
Who makes what · who's best at what · what's worth paying for
All three have their place. Only one is best at each thing.
The default everyone knows — and the one that's quietly fallen behind for pro work.
The pro's choice. Best thinking, best writing, best coding partner you can buy.
The swiss army knife. More modes than anyone else — video, audio, images, research.
Open, fast, relentless. Worth knowing — even if you don't use them daily.
The go-to open family. Shipping in pro-level quality. Runs locally. Strong at code and Chinese-English tasks.
The one that shook everyone. Frontier reasoning at a fraction of the cost. V3 and R1 are both serious.
Ridiculously long context. Great at reading through whole books and codebases and staying coherent.
Private. Free forever. Offline. No rate limits.
Google's open family. Punches way above its weight — a small, fast model that runs on a MacBook and is shockingly capable.
+ Qwen, DeepSeek distilled, Llama, Phi. The open ecosystem is exploding.
Desktop app. Browse a catalog of models, one-click download, chat in a clean window. The easiest on-ramp.
ollama run gemma — two commands, a model running. Lightweight, fast, scriptable.
Not all tools are created equal. Here's where each one lands.
Free, absurdly good. Feed it sources, get mastery.
Google's image gen. gemini.google.com → Create Image. Fast & free.
Emergent · Replit · Lovable. Describe an app, get an app. Real magic.
Pro video gen. Seedance 2 inside Runway is legitimately good.
Voice cloning & TTS — industry standard.
Full songs from a text prompt. Uncanny.
Gemini Deep Research or Claude Research already do the useful parts. Skip these.
Autonomous agent — big promises, murky payoff. Check before you trust.
AI with hands on your screen. Only as trustworthy as your ability to audit it.
Every subscription caps you. Claude's the tightest. The trick is knowing what burns tokens — and watching it.
A token is roughly ¾ of a word. One long answer can be 1,000+ tokens. Every message you send — and every message before it in the thread — gets counted again on the next turn.
A reasoning / thinking model can burn 5–10× more tokens than a cheaper one to answer the same question. Pick the smallest model that does the job. Reach for thinking only when you need the muscle.
In Claude: Settings → Usage. You can see which conversations burned the most tokens. Spot your expensive habits — long threads that should have started fresh, over-eager thinking mode, endless regeneration.
One idea. The whole game, rewritten.
LLM + goal + loop + tools
Code is the universal tool. Any task a human can describe — an agent can now write the software to do it. That's the unlock.
Three interfaces. Three ways an agent does things beyond just typing.
The agent drives your terminal. Reads files, runs commands, ships code. Exactly like a human engineer would.
A universal plug. Any app can expose its tools as an MCP server — the agent just plugs in. Notion, Figma, Linear, your own stuff.
The classic. Any service with an API — the agent can call it. Send email, pull data, trigger a workflow, move money.
Remember those movies? The alien taps into TV, the robot plugs into the internet — and in minutes it knows everything it needs. A Skill is that, but for your work: a small document the agent reads before acting. Hand it the way you do the job, and it just knows.
Where to get them, how to use them, what to watch for.
Anthropic publishes a library: anthropic/skills on GitHub. Plus a growing community of shared skills on the web.
A Skill can run code and touch your files. Treat installing one like installing any app — read it first, only pull from people you trust, and start with low-stakes tasks.
In Claude: Settings → Capabilities → Skills. In Claude Code: drop into .claude/skills/.
Claude ships with a Skill Creator skill — ask Claude itself to help you write a Skill for your work.
The next email you write might be drafted by your agent, read by their agent, replied to by theirs, and only surfaced to a human when there's a real decision. The humans stay in the loop — but the loop gets a lot faster.
For fifty years, software was made by a few, for the many. That just flipped. You can now describe a tool — in plain English — and have it. Not buy it. Not hire someone. Have it.
Tomorrow we take everything from Days 1 & 2 and map it onto your business. Know · Do · Show, one task at a time. You leave with a plan for Monday.