HM
3 AI Experiments
That Taught Me More About Myself
Than About the Tool
Hugo Melis · hugomelis.nl
HM
3 AI Experiments
That Taught Me More About Myself
Than About the Tool
Hugo Melis · hugomelis.nl
The Experiment
After Selling My Agency,
I Chose to Experiment Daily
Every day, one AI experiment. Documented openly.
Not tutorials. Not product reviews.
Personal confrontations with what AI changes about the way I work, think, and build.
Every experiment teaches me more about myself than about the tool.
— hugomelis.nl
30 experiments and counting · hugomelis.nl/experiments
Day 7
Running a Local Model to
Allow My Thoughts to Flow Freely
The Problem
Hugo records voice memos on walks — raw, unfiltered thinking. But sending those to cloud AI means his most private thoughts become someone else's data.
The Approach
Local Ollama models + OpenWhisper transcription. Nothing leaves the machine. Privacy becomes infrastructure, not policy.
Day 7 · Key Insight
Privacy is not about neighbors looking through the window. It is about how freely I allow my thoughts to wander.
Maximum experimentation, minimal exposure. Local models let Hugo think out loud without self-censoring — the trade-off between capability and surveillance dissolves.
Cloud
Local
Day 10
Troostwijk Auction Search Sucks.
Let's Fix That.
The Trigger
Misspelled auction listings ("mutliplank" instead of "multiplank") meant deals went unnoticed by other bidders.
The Build
Used Codex /plan mode + Docker. A bot that finds partial matches and misspellings on Troostwijk automatically.
The Result
Found auction deals invisible to normal search. A working prototype in about 2 hours of focused effort.
Day 10 · Key Insight
As AI gets better, my inspection threshold lowers. Leverage increases. Visibility decreases.
Before
Review every line of code.
Question each architecture decision.
Maintain full comprehension.
After
Skim the plan. Trust the structure.
Contain the environment, not the decisions.
Surrender comprehension for speed.
Day 19
Building My Own CRM:
Can the Vision Survive?
The Old Way
Rough idea → wireframes → designer back-and-forth → developer builds → realize it's not right → more revisions. Weeks pass. Budget runs out before the vision lands.
The AI Way
Skip wireframes entirely. Claude Cowork in Figma produces full-fidelity designs in a few iterations. The vision holder becomes the builder. No translation layer.
Day 19 · Key Insight
The cost of translating an idea into a working product has dropped to almost zero. We don't know what this means yet.
After years running an app agency, the gap between what Hugo envisioned and what got built was always wide. AI collapses that gap. The question shifts from "can I build it?" to "is this what I actually want?"
The Hidden Pattern
Every experiment follows the same loop:
01
Control
Set up the experiment. Define the question.
02
Confrontation
Run into what AI actually does vs. expectations.
03
Self-Knowledge
Learn something about how I work, not just the tool.
hugomelis.nl/experiments
