MVPs fail in two ways and you can usually spot which one from the backlog alone: too many features (because everything “feels needed”), and no testable hypothesis underneath them (so even shipping the whole thing wouldn’t tell you what you learned). These 12 prompts force a written “we believe X will do Y when given Z” before you scope anything, a brutal cut to 30% of the backlog, a Wizard-of-Oz or smoke-test alternative when that is cheaper than building, and a 6-week time-box with explicit kill / pivot / double-down criteria. Pair them with the jobs-to-be-done prompts to lock the underlying job before you cut.
TL;DR
- Write the hypothesis first (
We believe [persona] will [behavior] when given [value]; we'll know when [metric]), then scope only what tests it. - Cut the backlog to ≤30% of the original list. Mark every item “core test”, “support”, or “cut”, and demand a reason for each cut.
- Before you write code, check whether a smoke-test landing page, Wizard-of-Oz, or single-user MVP would answer the question faster and cheaper.
- Time-box to ~6 weeks with one metric and one written kill criterion, so a flat result triggers a decision instead of a quiet extension.
- These are reasoning tasks, not code generation. As of June 2026, paste them into Claude Opus 4.7, GPT-5.5 (Thinking), or Gemini 3.1 Pro and feed back real numbers from your traffic or interviews.
Which model to run these in
Scoping is a judgment task: the model is helping you argue down a backlog and pick a falsifiable metric, not write a feature. As of June 2026 the three frontier reasoning models all handle it well, and all three offer a usable free tier for a one-off scoping session:
| Model | Best slot for scoping | Free tier | Paid entry |
|---|---|---|---|
| Claude Opus 4.7 / Sonnet 4.6 | Long, critical “argue with my plan” sessions; 1M-token context | Sonnet 4.6 (limited) | Pro $20/mo |
| GPT-5.5 (Thinking) | Fast back-and-forth on the cut list | GPT-5.5, tight limits (US Free has ads) | Plus $20/mo |
| Gemini 3.1 Pro | Pasting a long backlog + research docs in one go (1M context) | Yes, in the app | Google AI Pro $19.99/mo |
Tip: turn on the “Thinking” / extended-reasoning mode for prompts 2, 9, and 12 (the brutal cut, the cost estimate, and the post-mortem). They reward step-by-step reasoning more than the others.
Best for
- Indie maker MVPs
- Internal MVPs at companies
- Pre-funded startup MVPs
- Validating a new feature in an existing product
- Proof-of-concept for B2B contracts
1. Hypothesis-first MVP scoping
I want to build an MVP for {idea}. Before scoping, write the testable hypothesis: "We believe {persona} will {behavior} when given {value}. We will know this is true when {metric}." Then scope only what tests this hypothesis.2. Brutal-cut MVP backlog
Below: my current MVP backlog of {N} features. For each, mark "core hypothesis test" / "support" / "cut". Justify cuts. The final MVP should have ≤30% of the original list.
{paste}3. Time-boxed MVP plan
For MVP {paste}, build a time-boxed plan: weeks 1-2 (build), 3 (launch), 4-5 (learn), 6 (kill / pivot / continue decision). Output: deliverables per week, the metric checked at week 6, the kill criterion.4. Manual / Wizard-of-Oz scoping
For idea {paste}, design a Wizard-of-Oz MVP that fakes the hardest tech with human labor. Output: what looks automated, what is manual, the costs, the limits of validation it provides.The Wizard-of-Oz pattern has a long track record: Zappos founder Nick Swinmurn validated online shoe demand by photographing shoes at a local store and buying them himself when orders came in, and IBM famously tested an early speech-to-text concept with a typist hidden in the next room. Use it when the expensive part to build is also the riskiest assumption.
5. No-code MVP scope
For idea {paste}, scope a no-code MVP using {tools — Bubble / Webflow / Airtable / Zapier}. Output: each feature → the no-code component, what is fragile, where it will break at scale.Rough pricing to plan the budget (as of June 2026): Bubble has a free plan, paid from ~$29/mo; Webflow has a low-cost entry Site plan (around the mid-teens per month) plus a higher CMS/Premium tier; Airtable free tier, paid from ~$10/user/mo; Zapier is task-priced and scales with volume, so it can creep from $20/mo into the hundreds. A common solo stack (Webflow + Airtable + Zapier or Make) lands around $60-100/mo at low usage and climbs from there as volume grows. Vendors reshuffle these tiers often, so confirm current numbers on each vendor’s pricing page before you commit.
6. MVP smoke-test landing page
Instead of building an MVP, design a smoke-test: a landing page with a fake-buy button. Output: page sections, the metric (sign-ups / clicks-to-buy / waitlist), the threshold that justifies building.7. Single-user MVP
For idea {paste}, scope a "single-user MVP": works for exactly 1 named user. Output: who that user is, what we hand-craft for them, what we learn, when we expand to 5 users.8. Build-only-the-spike
My MVP has 1 risky technical assumption: {paste}. Scope a tech-spike-only MVP that tests just that assumption. Strip everything else. Output the spike scope and the go/no-go threshold.9. MVP cost / time estimator
Below is my MVP scope. Estimate: build time (solo dev), build cost (with contractors), the single biggest unknown that could 2x the estimate, and the cheapest way to de-risk that unknown first.
{paste}10. MVP kill criteria
For MVP {paste}, define the kill criteria: which metric, which threshold, which timeframe. Output: the explicit "kill it" condition, the "pivot here" condition, the "double down" condition.11. MVP → product roadmap bridge
My MVP just validated. Below: what we learned. Design the 90-day post-MVP roadmap: what features deepen the validated value, what we deprioritize, what we say no to. Be explicit about scope discipline.
{paste}12. MVP post-mortem
My MVP failed validation. Below: what I shipped + the data. Help me extract: was the hypothesis wrong, was the execution wrong, was the audience wrong, was the metric wrong. Output a pivot vs kill recommendation.
{paste}Common mistakes
- No testable hypothesis behind the MVP, so shipping it teaches you nothing falsifiable.
- Building everything that “feels needed” instead of cutting to the one assumption that matters.
- No kill criteria, so a flat MVP gets quietly extended for another quarter.
- Treating the MVP as v1 of the product instead of a single experiment.
- Skipping cheaper alternatives (smoke-test landing page, Wizard-of-Oz, single-user) and going straight to build.
- Using vanity metrics (sign-ups, page views) instead of the behavior the hypothesis predicts.
FAQ
How small should an MVP actually be?
Small enough to test exactly one hypothesis and nothing else. The 30% rule in prompt 2 is a forcing function, not a law: if cutting to 30% still leaves features that do not touch the core assumption, cut further. A good test is whether you could run the experiment without the feature and still learn the same thing. If yes, cut it.
Smoke test, Wizard-of-Oz, or a real build, which do I pick?
Match the test to the riskiest assumption. If the risk is “does anyone want this?”, a smoke-test landing page (prompt 6) answers it for the cost of a domain. If the risk is “can the value be delivered at all?”, a Wizard-of-Oz or single-user MVP (prompts 4 and 7) delivers it manually so you skip the engineering. Only build real code when the risk is genuinely technical, which is what the spike prompt (8) isolates.
What’s a good kill criterion?
One metric, one threshold, one date, written before launch. “If fewer than 15% of waitlist sign-ups click buy by week 6, we kill it” beats “we’ll see how it goes.” Prompt 10 forces all three of kill, pivot, and double-down so a result in the middle still maps to a decision.
Can an AI model decide what to cut for me?
No, and you should not let it. The model is good at surfacing arguments for and against each item and at pressure-testing whether your metric is falsifiable, but it has no skin in the game and no read on your customers. Treat its cut list as a draft to argue with, then feed back real numbers from your interviews or analytics.
Do I need the $20 plan to use these?
Not for a one-off. As of June 2026 the free tiers of ChatGPT (GPT-5.5), Claude (Sonnet 4.6), and Gemini all handle a single scoping session. The paid plans help when you are iterating across a long backlog, pasting large research docs into the 1M-token context window, or running the heavier reasoning prompts repeatedly.