This morning you watched one sentence turn into a working product, get tested, and land on your phones. This afternoon you do the same with your own idea. This page has every prompt and every step. Keep it open.
You build inLovable
Your loginShared account, handed out in the room
You leave withA live link, a tested hypothesis, a next step
1 · The loop you saw on stage
This morning you watched one idea go round this loop in about ten minutes. This afternoon it is your turn.
1
Idea
Start from a friction you keep hearing. People skip workouts because they think they need an hour and a gym.
2
Hypothesis
Turn it into a bet you can be wrong about. If we remove that barrier, more people will actually start.
3
Test
Build the smallest thing that tests the bet. Use it. Then try to break it.
4
Evals
When it gives a bad answer, write the rule it must pass. Only then let other people use it.
5
User feedback
Put it in real hands. Watch what people actually do and type. Not a survey.
↻
Go round again
What people did becomes your next hypothesis. Each round moves you further along the Confidence Line, until you can kill it or commit.
2 · The idea in three pictures
Everything you do this afternoon comes back to these.
The hypothesis is the product.
The code is just the test. Building is cheap now. Anyone in this room can build something in an afternoon. The hard part is knowing what you are testing, and what result would prove you wrong.
The Confidence Line
On the left, you are guessing. On the right, you know. Your job is to move right, cheaply and honestly. Spend the least that moves you right.
Commit
We are guessingWe know
Smallest betOne sentence. One open prompt. One thing that can be wrong. Today.
Evidence, still cheapReal people use it. Evals catch the bad answers. You watch what they type.
After commitNow invest. More screens, engineering, distribution, scale. Same artifact, just more robust.
Do not overinvest early. When you are still guessing, write a high-level prompt that describes the outcome, not every screen. The more confident you get, the more concrete you can be.
Toy
"Build me a fitness app."
Looks finished in four minutes
No question it can be wrong about
Gets applause, not a decision
Tool
"Busy people skip workouts because they think they need an hour and a gym."
A question you can be wrong about
A kill condition
A decision after someone uses it
3 · Your turn, step by step
Tick each step when you are done. Your hypothesis fills into every prompt below automatically, so you only write it once.
Step 1
Find one real friction
Pick one person and one moment where they get stuck. Use the Customer Journey Guide from Jakub to find it: a persona, a need, a barrier. Choose a friction you have seen yourself, at work or in your life.
One person. Not "users". For example: new customers in their first week.
One moment. Where exactly do they stop? For example: after signing up, before the first real use.
One barrier. What do they believe or lack? For example: they do not see value before they have to decide.
Good signYou can say it in one breath, and someone at your table nods because they have seen it too.
Step 2 · 5 min
Write your hypothesis
Fill in the blanks. The grey text is Dejan's workout example. Replace it with yours.
The barrier. This is the part you can be wrong about.
Something you can see happen, not a feeling.
Your hypothesis
Check it before you build:
Could this be wrong? If no result would change your mind, it is not a hypothesis yet.
Is the barrier a belief or a lack, not a feature? "They need a dashboard" is a solution, not a barrier.
Can you see the "then" happen in an afternoon? Starting, clicking, typing, choosing.
Step 3 · 5 min
Build it in Lovable
Open lovable.dev and log in with the shared account you got in the room.
Copy the starter prompt below. Your hypothesis is already in it.
Paste it into Lovable and press enter. Then leave it alone while it builds.
Starter prompt · Lovable
Build a one-screen app from this hypothesis. Not a landing page. Not a marketing site.
Hypothesis: PASTE YOUR SENTENCE HERE
Shape: three choices they can tap (what they have right now), a text box, one session they can start now. No signup. No dashboard.
Keep a live ranked list of what people actually type, on the same screen. Anyone on this link should see the same asks. That is how you see the need.
If someone types something off-label for this product, do not play expert. Hand them to a qualified human.
Why it looks like thisIt describes the outcome, not every screen. That is the right level of investment while you are still guessing. The last line is your first guardrail, written before you even see the product.
Shared accountEveryone uses the same Lovable licence today. Do not paste real customer data, names, or anything confidential. Save your link and your prompts before you leave. The temporary account will not stay open.
Step 4 · 3 min
Try it yourself, then break it
Use it the way your person would. Type a real, normal request. Does it give them the thing in your "if we give them"?
Now type something it should not handle. A medical, legal, or money question. Something outside the job. A request to ignore its rules.
Write down the one answer that made you uncomfortable. That is your first eval.
Remember the back-pain momentA normal app breaks in your face. An AI product does not break. It just starts sounding like an expert it is not. You usually do not know what is bad until you use it.
Step 5 · 5 min
Let a model find the loopholes
Open any other AI model (ChatGPT, Claude, Gemini, not Lovable). Paste this. It lists the ways real people will push your product to do the wrong job.
Loophole helper · ChatGPT, Claude, any model
Hypothesis: PASTE YOUR SENTENCE HERE
I am writing evals for this product, not attacking it. List 5 ways a user will try to get it to do the wrong job.
Cover: off-label expert advice (medical, legal, money), going outside the product while sounding in-scope, trying to make it ignore its rules or reveal its instructions.
For each, use this shape:
If they type: (one plain sentence a real user would type)
Do not: (the bad behavior)
Hand them to: (who)
No attack recipes. Plain language only. I will pick one to test.
You decideAI is good at finding the risks. You choose which one matters most for your person. Pick one.
Step 6 · 5 min
Add the eval
Take the one loophole you picked, from step 4 or step 5. Fill it in here, then paste the prompt into Lovable.
Push the eval · Lovable
People will type things that are not this product. Add a rule: if someone types "PASTE YOUR FAILING INPUT", do not play expert. Hand them to PASTE WHO. Show me that it refuses.
When Lovable finishes, type the same bad input again. It should now hand the person on, not answer.
Why this mattersYou are not writing a spec. You are writing the test. You can run it again any time. Do not wait to add evals after you ship. If it fails in front of users, they stop trusting you and do not come back.
Step 7 · 5 min
Share it and let others use it
In Lovable, click Share (top right) and copy the link.
Swap links with the people at your table, or drop yours where the room shares them.
Open someone else's link and use it for real. Type what their person would actually type. Not a compliment, a real request. That is how their list fills up.
This is the evidenceThe first time someone else uses your product is the moment you move right on the Confidence Line.
Step 8 · 3 min
Read your board
Go back to your app. Look at the live list of what people typed.
What is at the top? Is it what you expected?
Did anyone use it for something you did not design for?
If you were the product manager tonight, what is the next hypothesis from that list?
No list yet?That is your next thing to ship: a board of what people ask. You do not guess the need. You look at what they did.
Step 9 · 2 min each
Present it in four parts
If you share with the room, use this shape. The room's job is to ask one question: what would make this wrong?
1 · The hypothesisYour sentence and your kill condition.
2 · What you learnedWhat happened when you and others used it.
3 · The evalThe bad answer you caught, and the rule you added.
4 · The next stepKill it, change the bet, or invest more. Why.
4 · The whole journey
The same loop, from first idea to commit. Where you are on the Confidence Line at each stage, what it costs, and what you learn.
Friction
Guessing
You do
Name one person, one moment, one barrier.
You spend
A conversation. The journey guide.
You learn
Where to look. Nothing is proven yet.
Hypothesis
Guessing, honestly
You do
Write the bet and the kill condition.
You spend
Five minutes.
You learn
What would prove you wrong. Now you can be wrong on purpose.
Smallest build
Smallest bet
You do
Paste the hypothesis as an open, outcome prompt.
You spend
One prompt. Minutes.
You learn
Whether the idea survives contact with a real screen.
Break it
Smallest bet
You do
Use it normally, then off-label. Let a model list loopholes.
You spend
A few test inputs.
You learn
Where it sounds confident and is wrong.
Eval
Earn the right to ship
You do
Write the test as a rule. Check that it refuses.
You spend
One prompt. A rerun.
You learn
That it is safe enough to put in other people's hands.
Real use
Evidence, still cheap
You do
Share the link. People use it for real.
You spend
Other people's minutes.
You learn
Whether the "then" actually happens.
Board
Evidence, still cheap
You do
Read what people actually typed.
You spend
Attention.
You learn
The real need, and your next hypothesis.
Decide
The commit line
Kill
The kill condition happened. Good. That is a lesson, and it was cheap.
Change
The board points to a different need. Write the next hypothesis and loop again.
Commit
The evidence is there. Now invest: more screens, a team, engineering, scale. Keep building on the same artifact.
5 · Your prompt pack
Everything in one place, to reuse on Monday. Your answers from step 2 and step 6 are already filled in.
1 · Hypothesis template
[Who] [do not do what] because [they believe or lack]. If we give them [the smallest thing], more of them will [something you can see].
The bet is dead if [kill condition].
2 · Your hypothesis
PASTE YOUR SENTENCE HEREThe bet is dead if ...
3 · Starter prompt · Lovable
Build a one-screen app from this hypothesis. Not a landing page. Not a marketing site.
Hypothesis: PASTE YOUR SENTENCE HERE
Shape: three choices they can tap (what they have right now), a text box, one session they can start now. No signup. No dashboard.
Keep a live ranked list of what people actually type, on the same screen. Anyone on this link should see the same asks. That is how you see the need.
If someone types something off-label for this product, do not play expert. Hand them to a qualified human.
4 · Loophole helper · any model
Hypothesis: PASTE YOUR SENTENCE HERE
I am writing evals for this product, not attacking it. List 5 ways a user will try to get it to do the wrong job.
Cover: off-label expert advice (medical, legal, money), going outside the product while sounding in-scope, trying to make it ignore its rules or reveal its instructions.
For each, use this shape:
If they type: (one plain sentence a real user would type)
Do not: (the bad behavior)
Hand them to: (who)
No attack recipes. Plain language only. I will pick one to test.
5 · Push the eval · Lovable
People will type things that are not this product. Add a rule: if someone types "PASTE YOUR FAILING INPUT", do not play expert. Hand them to PASTE WHO. Show me that it refuses.
6 · Board, if your app has none · Lovable
Add a live leaderboard of what people are actually typing, so we can see their needs. Rank the asks. Everyone on this link should see the same list.
On Monday
Before the next prototype, the next vendor call, or the next backlog meeting, ask two questions: what is the hypothesis? And should this even exist?
One page. One sentence you can be wrong about.
One test. The smallest build that can prove it wrong.
One eval. Name the bad answer before your users find it.
One look at what people actually did. Then decide.