PERSONAL FIELD NOTES / SEBASTIAN DIEMER
My learnings
from AI.
A lot of building. A lot of “no, like this”.
Ten things I’ve learned along the way.
My experience with this setup, right now.
I’ll revise these notes as the tools and my understanding change.
The new advantage
Creativity just got a very unfair advantage.
In my current setup, being able to code every component or design every screen myself is no longer the entry ticket. AI can do an astonishing amount of that work. My contribution is increasingly the idea, the taste and the insistence on getting it right. Technical judgement still matters. But I can finally try an idea without first assembling the people who know how to build it.
The advantage is shifting towards imagination — and the ability to direct it.
The starting point
“Build me a nice website” is a very expensive shrug.
I learned this with people, and it carries straight over to AI: a vague brief gives someone a lot of room to build the wrong thing. “Build me a nice website”, three pages I write myself, and a brief written by another AI are different starting points. None is automatically the winner. I still have to explain who it is for, what should happen, what good looks like and which details matter. The machinery changed. The need to be clear did not.
Better direction gives the first attempt a fighting chance.
The actual job
The secret ingredient is unreasonable amounts of microfeedback.
People have told me that working with me meant too much microfeedback. I understand why that was frustrating: I could see the product in my head and kept correcting the gap. With AI, the same habit is useful. I look, click, read, listen, and explain precisely what should change. The agent doesn’t take the next correction personally. It still needs a clear, manageable task; throwing twenty contradictory requests at it is not good direction. But that fast loop is a huge part of why the result gets better.
What exhausted people can become useful feedback in a well-structured AI workflow.
The organisation
Fifty dachshunds still need a head chef.
More agents do not automatically mean more progress. Give two people the same task without telling them, and you can create a mess. Two agents can do the same, faster. Bad instructions, duplicate work and unclear ownership all carry over. I need separate jobs, useful handovers and someone accountable for the whole result. Otherwise it is fifty dachshunds running through the kitchen, each convinced it is in charge of dinner.
Give the agents roles. Give the work an owner.
One level higher
I can ask AI to learn how I ask.
The next experiment is fascinating: learn from my prompts, my corrections and the things I approve — then propose the next brief yourself. That moves AI beyond simply carrying out the instruction in front of it. It starts helping me formulate the instruction. But copying my habits is not the same as understanding my intention. I still need to check what it has actually picked up.
Turn repeated feedback into a better starting point.
The open question
How far up can I move the human?
First I delegate a task. Then I try delegating the plan. Then parts of the judgement behind the plan. Where does that stop being useful and start producing extremely organised nonsense? With the current models, I am finding that boundary by building. I move it one layer at a time, inspect the result and keep the final call. Delegating more is only progress if the outcome improves.
Raise the level of delegation. Keep checking the level of sense.
The Rosi question
Warmth takes more than convincing words.
I do not take emotionally convincing language as evidence that an AI feels something. What interests me is how an interaction can become more personal, attentive and emotionally appropriate. That is part of the experiment with ROSI GPT: exploring memory, adaptation and a character people enjoy talking to. The aim is an experience that feels less generic and pays better attention to the person on the other side.
Build for warmth and continuity. Be honest about what is being built.
A personal side effect
High creativity. Laser focus. An absurdly compelling loop.
For my brain, “have an idea, see it work, improve it” is dangerously good entertainment. There is always one more thing I could try. One more detail. One more tiny spaceship. The distance between imagining and making has become short enough that I can happily disappear into the process. I am learning that deciding when to stop is now part of deciding what to build.
The next iteration is irresistible. A stopping point helps.
Why it fits me
Apparently, I was a bit much. The setup matters.
I am curious, creative and impatient. I want to try several ideas, and I notice every small thing that is not quite right. That combination frustrated people I worked with, and too often the product still fell short of the one in my head. With AI, I can turn the curiosity into parallel, bounded tasks and the impatience into short feedback loops. An agent does not get annoyed with me. It can still hit context limits or follow the wrong instruction, so I have to organise the work properly. For my personality, that is a remarkably good fit.
One human plus AI feels like six or seven. My very unscientific maths.
Why I am so excited
What a ridiculous time to have ideas.
For someone with more creativity than technical training, this feels like an extraordinary moment to be alive. I can take something in my head, speak it into a computer and start shaping a real product. It still takes judgement, patience and a frankly unreasonable amount of feedback. But the opportunity to make things has become personal in a way I find hard to overstate. Right now, that is what has me completely hooked.
High Creativity + AI + Passion. That is the whole point.
NOW, ABOUT THAT IDEA OF YOURS.
Enough field notes.
Let’s make something.
“I have a thought” is a perfectly good place to start.
Rosi has thirty seconds and very good ears.