Build real software with AI — and turn it into a business
16 modules · 177 min of material
The tool, and what to build with it
01 Claude Code Fundamentals Everyone's first bad session looks the same. Claude edits a file you didn't mean, forgets the instruction you gave twenty minutes ago, and burns usage on a task that should have been cheap. None of those are model failures — they are configuration and context failures, and each has a cause you can see and fix. This module makes Claude Code legible: what it loads, what it forgets, what it's allowed to touch, and how to diagnose it when it goes sideways. 12 min 02 AI Coding Mastery: One Real SaaS, End to End There is a gap between "Claude built me a login page" and "I have a product that takes money from strangers." The login page takes twenty minutes. The gap takes most people three months, because nobody shows them the middle: the schema that survives contact with real users, the webhook that fires at 3am, the container that runs the thing your serverless host refuses to run. This module closes that gap by building one product start to finish. Not twelve topics. One build, and 11 minHow you ask, and how many ask
03 Prompt Engineering for Code Most prompt engineering advice was written for chatbots and is actively wrong for coding agents. "Act as a senior developer" does nothing when the model already has your repository open, can run your test suite, and will read forty files if you let it. Claude Code doesn't need a persona. It needs to know which files matter, what it is not allowed to change, what shape the answer should take, and how it will know it succeeded. That is the whole craft, and it is learnable in an 11 min 04 Multi-Agent Development The fantasy is a boardroom: a CEO agent assigns work, the Architect argues with the Backend engineer, QA pushes back, and you watch your product assemble itself. That is not what happens, and believing it is why most people's multi-agent setups produce worse results than a single well-prompted session. Here is what actually happens: your main conversation is the only thing that can delegate, subagents cannot talk to each other, and each one starts with no memory of anything y 11 minGive it hands, then give it a P&L
05 MCP Masterclass You know the moment. You're debugging a production issue, and you paste the error from Sentry into Claude. Then the relevant row from your database. Then the Slack thread where someone described the symptom. Then the GitHub issue. You are a very expensive copy-paste pipeline, and every paste is a chance to lose context or leak a credential into a transcript. MCP exists to delete that job. When you find yourself copying data from another tool into chat, that tool should be a s 11 min 06 The AI Agent Business You can build a working agent in an afternoon. You still won't get paid for it. The gap between "my agent answers questions" and "a company wires me $4,000 every month" is not model quality — it's that nobody buys an agent. They buy a number going down (tickets, response time, no-shows) or a number going up (booked calls, qualified leads). This module covers the six agent businesses that actually close, what each one is worth, and the specific way each one dies in production. 11 minMake it repeatable, then make it unattended
07 The SaaS Factory Your last three products didn't die from bad code. They died at week six, when you finally reached the part you didn't know how to build — the OAuth refresh, the exactly-once publish, the custom-domain TLS — and discovered the interesting problem was buried under four weeks of login forms and CRUD screens you'd already written twice before. A factory fixes this by inverting the order: the boring 80% is a template you reuse, and the one genuinely hard part gets built on day on 11 min 08 Automation Masterclass The demo worked. You showed the client, everyone nodded, you set it to run every night. Three weeks later someone asks why the Tuesday report never arrived and you discover it stopped running on the 4th — silently, with no error, because the API returned a 502 once and nothing was listening. Or worse: it retried a payment webhook eleven times and charged a customer eleven times. The difference between an automation that demos and one that runs unattended for a year is about s 10 minWhich framework, and which cloud
09 AI Frameworks, Honestly Most people pick a framework before they know what they're building, then spend a month learning the framework instead of shipping. The truth nobody selling one will tell you: a large fraction of production "agents" are a bounded `for` loop around a model call with tool dispatch — about thirty lines — and wrapping that in an orchestration layer buys you dependencies, an upgrade treadmill, and a stack trace you can't read at 2am. Frameworks earn their keep on specific problems 11 min 10 Deployment Everyone learns platform selection the same expensive way: pick the one from the tutorial, build for three weeks, then discover your 4-minute video render can't run on a platform that kills requests at 300 seconds. Or you leave a GPU pod stopped and find a $100 storage bill for compute you never used. Platforms aren't better or worse than each other — they meter different things, and the meter is the product. This module maps workload shape to platform, shows the line items t 11 minGetting paid, and looking worth it
11 The AI Services Business You can build. That stopped being the bottleneck the day you finished Module 02. The reason most technical founders sit at zero revenue is that they price like an employee, they wait for clients to appear, and the first time somebody says "can't ChatGPT just do that?" they get defensive and discount 40% on the spot. This module is the commercial half of the job: where the first five clients come from, how to run a discovery call that ends in a signed scope, and how to put a n 11 min 12 Directing UI and UX in Words "Make it look better" is the most expensive prompt in this course. It produces a different result every time, none of them match the screen you built yesterday, and by screen six your app looks like it was designed by six people who never met. The problem isn't the model's taste — the model has plenty. The problem is that you handed it a feeling instead of a specification, so it picked a plausible answer from an enormous space of plausible answers. This module teaches you to 10 minDemand, and the discipline to serve it
13 Marketing That Isn't Generic "Write me a sales page for my SaaS" returns the average of every sales page ever written: *Transform Your Workflow. Unlock Powerful Insights. Join thousands of happy customers.* Grammatical, well-structured, converts at nothing — because it says exactly what your competitors say. That's a briefing failure, not a model failure: the model wrote the average because the average was all you gave it. This module covers what goes in the brief, how to structure an offer, and how to t 11 min 14 Shipping Safely When an Agent Writes the Code The dangerous failure mode isn't that Claude writes bad code. It's that Claude writes a lot of code, quickly, and you have no mechanical way to know something broke until a user emails you three days later. Velocity without feedback is just a faster path to an outage. This module installs the feedback: a CI pipeline that gates merges, logs you can actually query at 3am, alerts that fire on real problems, a rollback you've rehearsed, and a runbook so the person on call isn't i 11 minWhen it's not a demo anymore
15 Your AI App Has Real Users Now The demo cost $4 in credits and everyone loved it. Then 2,000 people started using it and three things happened in the same week: a $3,100 bill you can't explain, someone posted your full system prompt on X, and one user's 90-second request blocked everyone else's. None of these are model problems. They're the engineering that sits around the model, and none of it existed in the demo. This module installs it: the cost arithmetic, caching that actually pays, routing, rate limi 11 minThe harness as a library you import
16 The Claude Agent SDK Module 09 told you to write the plain bounded loop first and make any framework justify its cost. That still stands. But there's a case it didn't cover, because it isn't a framework: Claude Code is a harness — a loop, a tool set, a permission system, context management — and Anthropic ships that harness as a library you can import. So the question is not "SDK versus framework." It is narrower and far more answerable: what does the harness buy me that thirty lines of loop does 13 min