01
CLAUDE.md that agents actually read
What belongs in repo memory, what poisons it, and why “be helpful” costs you tokens.
From Loops to Orchestration —
No More Babysitting Agents
A hands-on video course on the context, verification and infrastructure that turn vibe coding into work you can trust.
No spam. One email when doors open.
Playing loads the video from Bunny (BunnyWay d.o.o., Slovenia), which receives your IP address, browser and referrer. See the privacy policy.
Hey — I'm Florian.
I ship software for a living: real customers, production systems, audits, the whole thing. And for the last year I've been doing it with coding agents — not as a party trick, but inside the same strict process everything else goes through.
Here's what I learned the hard way: the model was never the bottleneck. The platform around it is. Vibing is where everyone starts. This course is about what comes after — when you want the agent to actually be trusted with your codebase.
If you've felt that gap between a demo that dazzles and a change you'd dare to merge, this is for you.
— Florian van der Galiën
Lesson by lesson
Merge on proof, not on vibes
Some of it recorded, some of it still in production.
01
Context
What the agent knows.
02
Plan before code
Make the agent write the plan first, then review that instead of the diff.
03
MCP or CLI?
The call that saves 100k tokens a run: when a tool server earns its place and when the plain CLI is simply better.
04 · in production
Curating the loadout
Cutting tools until the agent stops guessing. Fewer, sharper, named.
05 · in production
Context that survives a long session
Compaction, handoffs, and what to write down before the window fills.
06 · to be scheduled
Being scheduled. What lands here comes from what people ask for.
02
Verification
How it checks itself.
01
The verification ladder
Linter, types, tests, preview URL, adversarial review. Where to start and when to climb.
02
Tests as eyes
Checks the agent can run itself, so “until green” finally means something.
03
Adversarial review
One agent implements, a second is paid to disagree — the pass that finds what a lenient self-grader misses.
04 · in production
Stop conditions
Teaching a loop when it's done, and what to do when it isn't.
05 · to be scheduled
Being scheduled. What lands here comes from what people ask for.
03
Infrastructure
What it runs on.
01
CI for a project of one
The pipeline that pays for itself on day one, on a hobby repo.
02
Preview deploys as ground truth
A real URL per change, so the agent checks its work in a browser instead of in its head.
03 · in production
An agent-friendly dev server
Background processes, logs it can read, and a stop command it can call.
04 · in production
Loops that run while you sleep
Queue the work overnight, review merge-ready branches with coffee.
05 · in production
Working under real constraints
What GDPR and a security audit actually change about an agent setup — and what they don't.
06 · to be scheduled
Being scheduled. What lands here comes from what people ask for.
The course is being built in the open. Titles, order and scope are subject to change.
Who's teaching
A senior dev who does this every day — not a course factory.
I'm a full-stack developer in a Web Dev & AI unit — Astro, Vue, Supabase, and a stack of coding agents inside a real, audited development lifecycle, where “just paste it into the chatbot” was never an option. I gave the “Beyond the Vibe” talk at the University of Bayreuth; enough people asked “how do I actually do this?” that I turned it into a course.
Husband, father of two, and genuinely convinced the job isn't shrinking — it's shifting from typing lines to designing the systems agents work in.
Claude CodeCodex CLIAstro · Vue · BunCI / preview deploysAudited SDLC
Get the first email when doors open.
No launch hype, no drip sequence. Just a heads-up when the course goes live.
Check your inbox.
If confirmation is needed, we'll send a 24-hour link to .
Questions I keep getting
Anything else? Email me — I answer every one. (yes, personally)
I already use Copilot / Cursor / Claude Code. Is this too basic?
If your workflow is “prompt, eyeball the diff, hope” — no. The course starts exactly where autocomplete and chat end: giving agents context, verification loops and infrastructure so they can work without you watching every keystroke.
I'm a non-technical founder. Can I follow along?
Yes, if you're willing to get your hands dirty. You won't need to write much code yourself — that's rather the point — but you'll set up real tools (a terminal, git, CI). I explain every step and why it matters, not just what to click.
Which tools does the course use?
Primarily Claude Code, with Codex CLI as a second perspective. But the course teaches the levers — context, verification, infrastructure — which transfer to whatever agent wins next year. Tool-specific parts are clearly marked.
How much time does it take?
Your call — everything is unlocked from day one and nothing expires. Lessons run 9 to 16 minutes; two or three a week is a comfortable pace. Every one ends with something running in your own repo, not a quiz — so the time invested compounds immediately.
What if the models change and this is outdated in six months?
Models change; the system around them doesn't. Better models make good context, verification and infrastructure more valuable, not less — they're what lets you hand more work over. That's also why the course keeps growing after you buy it — and why you pick how long you want that to go on: a year of new lessons, or every lesson I ever add.
What do I get, and what happens after a year?
Two options at checkout. Both unlock every lesson that exists the day you buy, and both keep it — nothing you already have access to ever expires. The difference is what comes later: one includes every new lesson published in the twelve months after your purchase, the other includes every lesson I add for as long as the course runs.