Building a proof of concept with Tim van den Bosch, 23 October
A prototype asks if they want it,
a POC asks if you can make it.
Prototype: no engine
POC: no car
Three things you can build,
each with its own question.
POC
Prototype
First product
It asks
Can I make it?
Do they want it?
Will they pay and keep using it?
Who sees it
You and your team
A few people you test with
Paying customers
What is real
Only the hard step
Only what people see
The whole core, every day
What you fake
Everything around it
Everything behind the screen
Almost nothing
Done when
Your number says yes or no
You know what to build
Never, you keep improving it
I learned it from
indie hackers on Twitter.
They invented the wheel years before AI: an idea live in a weekend, the numbers in public.
Headquarters, bunk 3.
Indie hackers prove each step
before they build the next.
POCCan I make it?
PrototypeDo they want it?
ProductDo they pay?
Pieter LevelsNomad List, 2014
A public spreadsheet: can strangers fill in the city data?
The sheet "quickly filled up".
The site, five weeks later.
Pieter LevelsInterior AI
Plays with Stable Diffusion: can it redesign a room?
"I quickly spun up an MVP."
A paid app.
Danny PostmaLanding page course, 2026
His own skill turns AI slop into a clean design.
A sales page for a course that does not exist yet.
Pre-orders at $29. The course opens in November.
Their POC is often not new code: a spreadsheet, or a skill they already use.
The best lessons in MAKE,
by Pieter Levels.
Use what you know
"Use what you already know and see how far you get with it."
Love your limits
Constraints "can be a giant advantage vs. people with lots of resources."
No code is fine
A luggage pickup startup without code. "It took about 30 minutes!"
The hard step must run
Fake the rest, never that step. "Make sure your MVP actually works."
Test it fast
"The best is to test your ideas as quickly as possible."
An afternoon, max
Levels gives a first version "max. one month". Your POC gets an afternoon.
For me, a POC is
proof that runs.
Before we built the POLE pension engine, I tested three things:
Which stack?Ruby in its own container.
Can it run in the EU?Yes, Bunny Amsterdam.
Is the maths exact?Yes, to the last decimal.
pole/rekenproef · July 2026
0.1 + 0.2 =
BigDecimal 0.3
float 0.30000000000000004
region AMS · EU
Floats drift at the 17th decimal, so the pension maths for POLE stays in Ruby.
My POCs:
one question each.
Feb 2025
Funda scraper
Can I pull business listings off Funda?
Worked after 9 minutes.
Mar 2025
SketchUp MCP
Can Claude draw in SketchUp?
Built in one day.
May 2026
Phone agent
Can an AI hold a natural Dutch phone call?
Yes, once the 7 s pause became 1 s.
Jul 2026
Rekenproef
Does the pension engine calculate exactly?
Exact, shown with one endpoint.
Aug 2026
Test API
Can we talk to the client's test platform?
No at first, then I found the right door.
Sep 2026
Oorapp
Can my Mac transcribe calls, without a bot?
Version 1.0 in two hours.
Oct 2026
POLE MCP
Can AI answer pension questions safely?
Live: the AI picks, the engine counts.
Oct 2026
Precedent
Can an architect's AI search built work?
A pilot on fake data in 45 minutes.
My archive holds 55 folders,
and most of them stopped.
Pieter Levels counts 4 hits in 70+ projects. When a POC answers no, close the folder and keep what you learned.
Every idea has one block
that holds it all up.
Pull out the logo and the tower stands. Pull out that one block and it falls. Your POC tests that block.
Spoiler: it's not the logo.
← This one.
Train the monkey first,
the pedestal is easy.
X is Google's moonshot lab, where the self-driving car began. Its puzzle: a monkey that recites Shakespeare on a pedestal. Boss Astro Teller: "spend zero time thinking about the pedestal."
The pedestal: logo, login, colours. Easy.
The monkey: your hard step.
Loon tested
the hardest step first.
Passed
Wi-Fi
Can a balloon send Wi-Fi to an antenna on the ground?
It worked, "and there were bets it wasn't going to".
Passed
Phones
Can it talk to phones directly, without the antenna?
Yes.
Passed
Speed
Is it fast enough for real internet?
From under 1 to 15 megabits per second.
Passed
Steering
Can a balloon sail the winds to where it's needed?
A balloon inside a balloon.
Passed
Lifetime
Can it stay up 100 days, at under 5% of the usual cost?
"Yes. In the end." After many busted balloons.
Killed
Cargo airship
Can X test it cheaply?
No: $200 million for the first data point.
Project Loon wanted internet from balloons in the sky. Astro Teller, head of X, TED 2016: "Run at all the hardest parts of the problem first."
Thomas, where are you?
Thomas would say:
start with what you have.
Bird in hand
Who are you, what do you know, who do you know? Build your POC with that.
Affordable loss
Pick a test you can afford to lose. An afternoon, not your savings.
Crazy quilt
Can't make it yourself? A partner who says "we can do that" is your proof.
Effectuation, from Saras Sarasvathy's research on expert entrepreneurs.
Which block sinks your idea
if it fails?
Matching app
Does your code pick the match a human would pick?
Food box
Can you buy the food at this price?
Delivery service
Can you deliver within 24 hours?
Build that step
and fake the rest.
Skip the design, the login and the extras. Only the step you test has to be real.
Only this door is real.
Set your number
before you shoot.
Write it down first, like "8 of 10 right". Pick it afterwards and every test looks like a win.
Don't be her.
You can build it yourself
in an afternoon with Claude.
Trust your number, not your feeling: in 2025, developers with AI felt 20% faster and were 19% slower.
Bread: done. App: almost.
Give me your riskiest step
and we build it now.
Watch the four steps you do next: find the block, write the number, build, check.
Live demo, 20 minutes. What could go wrong?
Your turn:
prove your hardest step.
Too big to build today? Then the smaller step you found is your result.
It works on my machine!
30 minutes
to prove one step.
Find the block
5:00click or press T
Brainstorm: what do you already have, and what blocks you?
Set the number
5:00click or press T
What result counts as proof? Write the number down.
Build it
15:00click or press T
Build it in any tool, or act it out by hand. It has to really happen.
Show it
5:00click or press T
Three volunteers show their number. The rest in the coaching hour.
A failed POC this month
beats a failed product next year.
Run your test for real this week and bring the number next Friday.
Test 1 went great.
Follow the people
who build this way in public.
Pieter Levels
@levelsio
Started Nomad List as a spreadsheet and gave every startup one month.
Marc Lou
@marclou
Caps every build at 14 days: build one feature, ship fast, kill faster.
Danny Postma
@dannypostma
Builds an "absolute alpha" first to check if it can be done.
Jon Yongfook
@yongfook
Did his own 12 in 12 with one-month MVPs, then built Bannerbear.
For the thinking behind it, follow Arvid Kahl (@arvidkahl) and Rob Walling (@robwalling).