I use a 90-minute sprint to decide whether an idea deserves more time. Ninety minutes is my operating constraint, not a universal promise. The output is not proof that a company will work. It is a small package of evidence: a specific customer, their language, a falsifiable offer, and a handful of real conversations.
The goal is to make being wrong cheap.
1. Write the kill test
Before research, I write down what would make me stop. Examples: I cannot find ten reachable buyers with the problem; existing workarounds cost less than the switching pain; nobody will agree to a conversation; or the required delivery model violates a constraint I will not accept.
This prevents the agent from turning every weak signal into encouragement.
2. Choose one beachhead customer
“Small businesses” is not a customer. I choose one person I can describe and reach: their role, business type, current workflow, triggering event, and the consequence of doing nothing.
A narrow beachhead does not mean the total market is small. It means the first message can sound like it was written for someone.
3. Mine public customer language
Customer interviews are valuable, but they are not always the first available evidence. I start with public material buyers already wrote: support communities, reviews, job posts, forum questions, product comparisons, and comments.
I collect exact phrases along with links and dates. Then I group them into problems, desired outcomes, objections, and existing alternatives. I do not let the model fabricate quotes or pretend public research replaces consent-based customer discovery.
4. Build a one-page kill list
Instead of a large business plan, I create one page:
- customer and triggering event;
- painful current workflow;
- proposed outcome;
- alternative solutions;
- why the customer might ignore me;
- delivery risks;
- evidence that would advance or kill the idea.
The page is a decision tool, not investor theater.
5. Sell a service before building software
When possible, I test the result manually. If the idea is an analytics product, I may deliver the analysis as a service. If it is automation, I may run the workflow with human review behind the scenes.
This reveals the actual inputs, exceptions, and trust requirements. It also avoids hiding a weak offer behind a polished interface. I disclose what is manual and never imply a prototype is more autonomous than it is.
6. Launch with ten personal messages
I skip the generic announcement post. I send a small number of relevant, personal messages that explain why I am contacting that person, what I observed, and the low-friction next step.
The purpose is learning. I track replies, objections, calls, and the language people use. No scraped blast list. No fake urgency. Ten thoughtful conversations teach me more than a large vanity impression count.
7. Anchor price to value, then test it
Cost-plus pricing tells me what I need. Value research tells me what the outcome may be worth to the customer. I estimate the economic or operational value, compare alternatives, and state a price that makes the tradeoff clear.
Then I ask. A pricing hypothesis without a buying conversation is still a guess.
What I keep after the sprint
At the end I want a dated evidence folder, not a motivational score. It contains the customer definition, source links, language clusters, kill criteria, offer, outreach copy, responses, risks, and the next smallest experiment.
Sometimes the answer is stop. That is a successful validation sprint. It returned time before the idea could consume a quarter.