Claude Code is the tool we actually build in. This is the real course, the moves from our own sessions, not the version posted for engagement. It is visual, you can touch it, and it updates as the tools shift.
The shift that changes everything: you are not asking questions and copying answers. You are directing a capable builder, setting the context, the constraints, and the taste, then catching it when it is wrong. Everything below is how you direct it well.
You are not stuck with one brain. Three models cover almost everything, and knowing which to reach for is the cheapest upgrade to your output. Tap a task and watch the right one light up.
The deepest reasoning in the family. Reach for it on hard architecture, gnarly debugging, and planning that has to hold a lot in its head at once.
Costs the most, so save it for the work that earns it. Turn on /fast for the same Opus with quicker output.
The balanced default. It handles most coding cleanly and quickly, at a fraction of the cost. Start here, and step up only when a task fights back.
If you are not sure which to use, this is the answer nine times out of ten.
The quickest and cheapest. Built for volume: mechanical edits, quick lookups, and the noisy background work you hand to subagents.
Speed and price make it perfect for running many small jobs at once.
Switch any time with /model. Fable 5 rounds out the Claude 5 family too. The rule of thumb: default to Sonnet, drop to Haiku for volume, step up to Opus for the hard parts.
A skill is a small file of instructions and context you hand it one time, so it works the way you work before you ask for anything. It loads only when it is relevant, so it costs you nothing until it is needed.
--- name: brand-voice description: Our house voice. Use for anything client facing. --- Write grounded, warm, and refined. Substance over noise. Never use em dashes. Recast with commas, colons, or periods. Read it aloud: if it sounds like a LinkedIn post, rewrite.
The description is the whole trick. Claude reads it at the start of a session and pulls the skill in on its own the moment a task fits, without you having to remember it is there.
A paragraph of setup you type over and over becomes a single command. Some ship with Claude Code, and the real power is writing your own. Type in the palette to filter, and open any one to see what it does.
The leap from fast to unfair is delegation. Instead of one long conversation carrying everything, you spin off focused workers, script the repeatable jobs, and wire in rules that run on their own.
Hand a whole task to a focused agent. It works in its own context and returns just the result, so your main thread stays clean. Run several at once.
For work at scale, a script orchestrates many agents in the background: a big migration, a repo wide audit, research cross checked from every angle.
Deterministic rules that fire on their own, like formatting after every edit or blocking changes to a protected file. They run whether or not the model thinks to.
MCP connects Claude to the apps you already use, so it can read and act on them directly instead of you copying data into the chat. GitHub, your database, Figma, Notion, the whole shelf.
Read a live database and act on it
Open and comment on issues in GitHub
Pull a design straight from Figma
Draft into the docs you already keep
Tool details load only when Claude actually reaches for them, so a shelf full of connections does not weigh down every message.
Tokens cost money and time costs more. Everything you have seen so far pays off here. Flip the habits on and watch the spend fall. This is the difference between people who love the tool and people who quietly give up on it.
Every default, nothing tuned. This is the money pit.
Illustrative, not a benchmark. The point is the shape: a handful of habits, stacked, is the whole game.
Context is the working memory of a session. Two files and two commands are most of the game.
Standing instructions loaded every session and kept through a compact: your conventions, your stack, how you like things done. Keep it tight.
Notes Claude keeps for itself across sessions on a project, so hard won context is not lost when a conversation ends.
At the seam between tasks, trade the history for a summary and keep the room to think. Add a focus to steer what stays.
When the next task shares nothing with the last, wipe it clean so nothing stale bleeds into fresh work.
No. The gap between an idea and a working thing has never been lower. If you can say clearly what you want, with the right context and constraints, you can direct real work. The taste is yours, the model supplies the speed.
Default to Sonnet 5. Drop to Haiku 4.5 for fast, high volume, mechanical work. Step up to Opus 4.8 for hard architecture, subtle debugging, and planning that has to hold a lot at once. Switch any time with /model.
Be specific up front. Most wasted tokens are corrections, and corrections come from a vague first ask. After that: send search to subagents, plan before you build, and compact at the seams between tasks.
Compact when the next task builds on the last and you want the thread to keep the gist. Clear when the next task is unrelated and you want a clean slate with nothing carried over.
No, and that is the point. It is a living course. It updates as the tools shift, and more lands every month. Bookmark it.
Every move on this page, condensed to one printable sheet, plus the skill file and prompt pack we build with. Drop your email and it opens right here.
Made it to the end?
Mark the course complete and claim your credits for finishing.
The creative who learns to build right now has an unfair advantage for years.
This is a living course. Leave your email and every new chapter, plus the prompts and skill files we actually use, lands in your inbox.