· 4 min read · by Thorge Dammann
How I Ship Software in Days, Not Quarters (With AI Agents)
No magic, just a process. The 5 rules I use as an AI-powered developer to ship working software fast with coding agents, without creating a mess.
TL;DR
- ✦Pick one use case. Not ten.
- ✦Use real data from day one. Fake data hides the real problems.
- ✦Let AI agents write code, but in small, verified steps.
- ✦Show working software every week. Not slides.
- ✦Decide with numbers, then scale, hand over, or kill it cheaply.
“How are you so fast?”
I get this question a lot.
Here’s the honest answer. No magic. Just 5 rules. 👇
Rule 1: One use case. Not ten.
Every AI project starts with a list of 15 ideas.
We pick one. The one that’s valuable, boring and measurable.
Boring is good. Boring means someone does it every day, and will notice when it’s gone.
Rule 2: Real data from day one
Demo data makes everything look great.
Real data shows you the messy PDFs, the missing fields, the edge cases. That’s where AI projects actually succeed or fail.
So we start there.
Rule 3: AI writes code. In small steps.
Yes, I use AI coding agents. A lot.
But never with “build me the app” and hope for the best.
Instead:
- Write a short spec
- Let the agent do one change
- Validate it. Commit it.
- Fresh context. Next step.
I even built a small supervisor to keep agents on that track.
The result: the speed of AI, the reliability of boring engineering.
Rule 4: Working software every week
Not a status report. Not a slide.
Something you can click.
Weekly demos kill bad ideas early and make good ones better. And they keep everyone excited. 🚀
Rule 5: Decide with numbers
At the end, you know:
- How good is it? (quality)
- What does it cost per request? (money)
- How fast is it? (latency)
- Do people use it? (adoption)
Then you decide: scale it, hand it over, or kill it cheaply.
All three are wins. The only loss is spending a year on a slide deck.
The bottom line
Speed isn’t about working more hours.
It’s about a small scope, a tight loop, and tools that do the typing for you.
Got an idea sitting in a backlog somewhere? Let’s talk.
FAQ
What is agentic coding?
+
Using AI coding agents (like Claude Code) to write, test and refactor code in small, supervised steps. The developer sets direction, reviews and verifies. The agent does a lot of the typing.
Isn't fast the same as sloppy?
+
Only if you skip the verification. Small steps, specs and tests are what make speed sustainable. Fast and sloppy just moves the work to later.
How long does a typical AI prototype take?
+
A first working version usually exists within days. A full prototype sprint with real users and honest numbers takes 4–6 weeks.