
The Claude Code Wars
Stefan Munz, Agentic Coding Meetup Hamburg, July 2025
Why are vacuum robots fighting here? Wait till the end...

But First: Talent War
Anthropic landed a hit with Claude Code.
What's so special about it?
The combination of excellent models optimized for coding with a clever terminal UI hit a sweet spot.

Cursor looked into the VC playbook and did what any good vc backed competitor would do: Poach the top talent.
It's an asset for you, but most importantly it hurts your opponent.

Meta: "Hold My Beer!"
200.000.000$ for one manager
Meta realized they have nothing to compete in this market for now, no model and no slick terminal UI
So they started handing out NBA salaries for individual top talents
first for OpenAI employees and then for Apple employees.
these talents earn more than most of the top CEOs in the USA.

Google: "💸💸💸"
2.400.000.000 $ for top talent & licensing
How can we get into the game, too?
Throw 2.4 billion usd at windsurf for some top talent and software licensing

Why?
I'm on the 200 USD plan from Anthropic
Neat little tool Ccusage lets you see your estimated token spend.
What would you have spent?
My last month is more than 700 USD
Totally worth it.
✅ The AI Hype Found Its First Business Case: Agentic Coding
🕵️♀️ The Real Claude Code Wars
Team YOLO vs Tight Leashers
This is a meetup about agentic coding, not about big tech strategies.
The real war is about how to get the most out of your Claude Code subscription
Peter Steinberger is a famous iOS & Mac developer
He shared a 3 hour video about how he vibe codes
Very watchable, but already a bit outdated
He made me buy the Max Plan
Andrej famous tech leader, worked at Tesla & OpenAI
He coined the term vibecoding, gave a presentation at YCombinator
I recommend watching this video, but I don't recommend his way of working with coding agents, part of the tight leashers.

🐍 The Tight Leashers
Read it out.
This means every loc is owned by a human

TEAM YOLO 🤷♀️
Yolo people make good backups of their computer, then give full control to the agent.
Let the agent run autonomously, human only reviews bigger chunks of code, if at all

The War Zone: Social Media
People on both fronts are fighting vigorously.
What Should We Do?

8 Agentic Sessions in Parallel 🤯
It's a pretty extreme world right now
I don't recommend this, it induced serious headaches for me
Constant Change of Context, feels more like babysitting toddlers than coding
Weird in-between step: I want massive output, but don't want to let go of control
The Return Of The Software Spec

ridiculously easy to go from pet project to open source
Written in Go, with no prior experience in Go
proper read-me, code of conduct, mit license
unit tests & linting. Most time spent on aligning versions between my setup and CI with GitHub Actions
It even has an image

I did not touch a single line of code in here, just the readme
The last project I vibe coded without a spec on a Friday evening

project-manager takes 3 files as an input
There's an example in the repository

Expand Your Specification with Gemini 2.5 Pro
The most important thing.
Take your time for this, write it with ai.
Let another ai review this
Do another round.

Gemini is good at slicing, too!
Then, slice the spec into tickets.
A refinement meeting

The prompt is mostly useless
Don't worry too much about the prompt.

💵🔥
Time to burn credits
Project manager is a simple for loop
For each ticket one agent is started in YOLO mode
One agent at a time, because usually the tickets build up on each other
This prolongs my coding days and output
Speccing and QA are my new bottlenecks right now

Focus on the Result, not how your Agent get's there!
This is the TLDR of this talk.
Read it out loud
Brings us back to vacuum robots, where it's the same.
I am really not alone on this.
Very good talk, making the point that our constitutions are legal specs
AWS just yesterday realized a new coding agent
Emphasis on a spec phase before coding starts
YOLO 🚀
But most importantly, try it for yourself!
Thanks you very much for listening!