Today is January 1 — the best time to recalibrate how you work.
I collected all available materials on how Y Combinator recommends forming hypotheses. This article is a distilled summary. In the sources, you’ll find summaries and full transcripts of videos or articles — you can upload them into your own neural networks.
Custom GPT
I uploaded all these materials into ChatGPT — you can test your hypotheses right now: YC Hypothesis Evaluator
What a Startup Hypothesis Consists Of
At YC, an idea is treated as a set of hypotheses:
Problem Hypothesis
There is a specific group of people who regularly or acutely experience a certain pain. They already solve it somehow and are dissatisfied with existing solutions.
A good problem usually has several characteristics: it is urgent; occurs regularly; costs time, money, or emotional energy; and cannot be ignored.
Solution Hypothesis
Your product, even in its simplest form, alleviates the pain better than what people currently use.
The MVP should be as simple as possible — something you can build in days or weeks. It’s important not to fall in love with the product: it will almost certainly work poorly, but it will give you data on what to change next.
Insight Hypothesis
You have an unfair advantage — a reason why you specifically can do this. Most often this is founder–market fit: you’ve personally experienced the problem, worked in the industry, or have access to early users or data — and therefore see what others don’t.
Market Hypothesis
The problem has a sufficiently large or growing market. Often the “why now” is explained by changes in the world — technology, behavior shifts, new infrastructure — that create new needs and drive demand.
A practical test from Paul Graham’s essays: who will use this right now — enough to accept a rough version from an unknown team? If you can’t name such people, the hypothesis is likely weak.
Hypothesis Template
- problem: [Who] [how often] needs [pain]. They use [current solution] and are unhappy with it.
- solution: If we provide [MVP], then [users] will [behavior change], and this will solve [pain] better. We will validate this via [metric].
- insight: Now is the right time because of [market change]. Others don’t do this because of [barrier]. We can execute because [why us].
Sources
How to Get and Test Startup Ideas – Michael Seibel
The initial idea can radically transform (justin.tv → Twitch). Start with the problem. Keep a “problem book.” Personal connection to the problem. Brainstorm with friends (cofounder fit).
Unique insight/angle/experience. Study existing/failed solutions. Build MVP fast (days/weeks), don’t fall in love. First users are hand-picked, with a strict filter. Validation: does it solve the problem for a specific segment, not “how many want it.”
How to Get and Evaluate Startup Ideas | Startup School – Jared Friedman
Start with a real, specific problem (not CISP and not a tar pit; google/analyze “why it’s hard”). Evaluate: founder-market fit, market, pain severity, insight vs competitors, recent change/proxies, scale, idea space. “Pros disguised as cons”: hard to start, boring, competitors. Best test — just launch.
How to Evaluate Startup Ideas – Kevin Hale
Idea = hypothesis of fast growth: problem → solution → insight (unfair advantage). Good problem: mass/growing/urgent/expensive/mandatory/frequent + needs triggers (Fogg). Not CISP. Unfair advantage: founder, market, 10× product, free distribution/word-of-mouth, network effects.
How to Get Startup Ideas – Paul Graham
https://paulgraham.com/startupideas.html
Don’t “think up ideas” — notice problems, ideally your own: made-up ideas sound plausible but get zero users. A good idea = you want it yourself, can build it, others don’t see the value.
Look for demand like a “well”: a small segment wants it very badly right now (who uses a broken v1?). The “path out of the niche” is often invisible — become a prepared mind: live in the future, build what’s missing/interesting, turn off schlep/unsexy filters. Don’t fear competitors: crowded market = demand; you need a thesis on what incumbents are missing.
Ideas for Startups – Paul Graham
https://paulgraham.com/ideas.html
A startup idea is not “a million dollars,” but a question or starting point that, by breaking, leads to the real idea; phrase it as a question, start with a slice of the problem and expand.
The best ideas are born “upwind”: new technologies + the right friends/cofounders (together-alone-together). Sources are annoying problems that “should be solvable”; make things simpler/cheaper/more convenient, often by redefining the problem. Often success = acquisition → build what a few buyers will want. The most common path is building for yourself and “fun hacks” with friends.
What Startups Are Really Like – Paul Graham
https://paulgraham.com/really.html
A startup is not a “job,” but everything at 10×: cofounders are critical (character/commitment; relationships must be maintained), the company consumes your life, the swings are brutal. Persistence carries you through; everything takes 2–3× longer (especially deals), so think in 3–5+ year horizons; start minimal, launch fast, talk to users, and pivot.
There is no “killer feature” — success = many small iterations; competitors are overrated, acquiring users is hard; deals fall apart, investors are often irrational, luck is a huge multiplier, community helps.
Interesting Discussion on the YC Forum
https://news.ycombinator.com/item?id=8167340
A thread about validating business ideas: positive feedback/early “customers” do not yet mean demand — they’re often not ready to pay. Collecting emails on a landing page is also a weak signal: signups don’t convert into usage/money, so validation should be done via real commitments (payment/pre-order/time).
YC Requests for Startups (RFS) — Example Product Hypotheses
https://www.ycombinator.com/rfs
A bottleneck of the AI boom is the shortage of skilled trades (electricians, HVAC, welders) for data centers and factories; the AI Action Plan launches and funds rapid reskilling. Hypothesis: a vocational school for the “AI economy” — personalized programs + hands-on practice with an AI mentor (voice/camera) and/or AR/VR simulators.
Money: employers pay for job-ready people; an AI mentor scales, unlike human tutors (the core problem of bootcamps).