The durable layer beneath the tools

A small-group program for people who want to build real web apps with Claude Code — and walk away with skill that survives the next model release, not just familiarity with this month's CLI.

Course structure and what you get

8 hours across 4 weeks. Designed so you walk away with something real, not just a certificate.

What you walk away with

• 01
  Your own real web appDeployed and live on the internet — actually usable
• 02
  A way of thinking that designs firstKnow what you're building and why, before writing a line of code
• 03
  An agent harness you can reuseSet it up once, use it on every project that comes after
• 04
  Real understanding of why AI behavesAnd the levers to make it behave the way you need
• 05
  The testing and deployment foundationsThe basics professional teams rely on — so your work lasts
• 06
  A way of working that does not expireWhen the next tool launches, you absorb it instead of starting over

Who this is for

• A
  You can use a computer comfortably — open files, run commands
• B
  You have some IT background and skill, but you don't need prior programming experience
• C
  You can commit a laptop, internet, and 2 hours per week
• D
  You have a real project in mind — we'll shape it together in the intro session, ready for real-world use

We meet before the course officially starts, so the 8 hours inside the course are spent on the highest-value material — not on setup or scoping.

What we cover in this session

• 01
  Get to know youYour background, current work, and what you want from the course
• 02
  Discuss the real projectThe one you'll carry as a through-line for all 4 weeks
• 03
  Walk through the course outlineWhat each week unlocks for you, specifically
• 04
  Check machine and accountsEverything configured before week 1 begins

What you leave with

• →
  Clear picture of whether this course is right for you
• →
  A real project, scoped and ready to be the through-line
• →
  A configured machine — no setup time wasted on day one

Some people use AI and their work ships faster. Others use the same tools and watch their projects fall apart. The difference isn't coding skill — it's a clear mental model of how AI works, how a codebase is laid out, and which knobs you can turn. Week 1 builds that whole foundation — from inside the model to the real software pipeline.

What we cover this week

• 01
  Why fundamentals matterComprehension Debt, the responsibility AI can't hold, and why you still must understand what's underneath
• 02
  How LLMs actually workthe ages of AI, the transformer, why answers vary, the context window, the agentic loop, and the onion architecture
• 03
  Read a codebaseproject structure, modularized MVC, naming, and the convention.md + design.md that steer AI across the whole team
• 04
  The 5-stage fundamental pipelinespec → data → ui/ux → test → deploy with each stage's real tool, plus extracting the design from a real app (the FlowAccount case)
• 05
  Control the outputModel + Effort (the builder's ceiling), the Prompt × Model × Context circles, why the CLI, and where to get your signal
• 06
  Workshopsource control as your time machine: first save point, first commit, GitHub + SSH key, and seeing the timeline in Sourcetree

Most people use AI by asking it to "do the thing" — and the result changes every time. This week you get agentic coding's full control system — instruction files, spec packs, hooks, rules, and skills — installed together, so output becomes steady and verifiable.

What we cover this week

• 01
  The three agentic toolsClaude Code · Codex · Gemini, all steered by files + the CLAUDE.md load hierarchy straight from the Claude team
• 02
  Two example spec packsmini-account (exact money) and villa-booking (non-colliding time), with xlsx/csv twins and the SQL-extraction prompt
• 03
  Anatomy of the control systemCLAUDE.md, settings.json (permissions/deny), and the agent's full tool catalog
• 04
  The full hook familyPreToolUse · PostToolUse (format + unit-test/lint) · Stop (defining done) + the circuit breaker and voice alerts
• 05
  Rules · Skills · doc routingscope-loaded rules, skills as packaged methods, MD routing, and the command/skill/agent comparison across all three platforms
• 06
  Workshopinstall the full control kit onto the villa-booking app: CLAUDE.md + mastermind.md + rules + hooks + a command-skill

One agent can write a feature. Building a system takes several agents working together — and a codebase AI can still navigate six months from now. This week we turn last week's single agent into a team, with tests that actually catch bugs.

What we cover this week

• 01
  The skill forge: _scratchthe workbench where thinking happens before code, and how to crystallize what works into reusable files
• 02
  Learn from the real thingMatt Pocock's 22-skill kit and Andrej Karpathy's 4 principles + nano repos, all pullable to read
• 03
  Real commands from the kit/big3 (three-perspective council), /scratch, /ssot-squad — how to shape repeatable skills/commands
• 04
  Subagent vs Agent Teamwhat differs, when to use which + subagent anatomy and the /teamx1-v2 review cycle
• 05
  Official Agent Teams + Plan ModeAnthropic's official feature (how to enable) and planning before editing
• 06
  The anatomy of a reliable agentwhat an agent really is, output control, small blocks, tools, and the agentic loop

Something running on localhost is not a product. The final week takes your project to a real domain, closes the holes you can close, and sets up systems that keep working for you after the course ends.

What we cover this week

• 01
  Deterministic spec-driven work/addtask + /cycle + task.csv, drafting specs in scratch then /update-truth, and output-shaping skills
• 02
  The real test kitunit/normal/edge, Playwright (2 workers) · Newman · curl, and where tests live: hook/plan/skill
• 03
  Test tools — web and mobilePlaywright, Newman, curl, hands-on simulator debug, Patrol
• 04
  Security with the right timingOWASP Top 10, free scanners (ZAP · SonarQube), and the day-one essentials (keys/ports/rate-limit)
• 05
  Deploy by script, not by faithdeploy.sh / release.sh, a /command as the trigger, and Jarvis-style automation (claude -p + cron)
• 06
  Workshoptake your project to real production, with a test suite and a deploy script