Guide

Strategy and planning

Types of MVP

Concierge, Wizard of Oz, single-feature wedge, or glued tools — pick the cheapest honest signal.

49 of 128 · 2 min read

Once you have a riskiest assumption, pick the cheapest method that produces honest signal. The type of MVP is a choice about what you must observe, not a choice about architecture.

Four methods

Concierge

You perform the service manually for users, as if software existed. High touch, deep learning, does not scale — that is the point. Food delivery startups that took orders by phone; onboarding consultants who configure the product for each design partner.

Wizard of Oz

The user thinks the product is automated; humans operate behind the curtain. Classic example: early testing of recommendation or AI features where a person curates results. Great for testing the experience before the model or pipeline exists. Do not deceive about safety-critical behavior; be ready to disclose in research settings.

Single-feature / narrow wedge

Ship one job end-to-end for one segment. Dropbox's famous early demo video tested demand for sync before the hard multi-platform client was done. Instagram launched as a fast photo-sharing app, not a full social network with video, shopping and messaging.

Piecemeal

Glue existing tools (Typeform + Airtable + Zapier + Stripe) into a workflow. Ugly internals; real payments and real usage. Replace pieces with custom software only where the glue breaks.

Paul Graham's essay Do Things that Don't Scale is the same idea in founder language: recruit users one by one, manually onboard, even perform the service by hand. Stripe's "Collison installation" — when someone said yes, the founders set up Stripe on their laptop immediately. Airbnb founders photographed listings themselves. Unscalable work teaches you what to automate. Automating first teaches you to scale the wrong thing.

Trade-off

Speed of learning vs brand risk. Concierge and Wizard of Oz learn fastest but can confuse the market if you over-advertise. Single-feature products protect brand quality but take longer. For pre–product-market-fit, bias to learning speed with a small, invited audience. For trusted brands adding a line, bias to a polished thin slice for a beta cohort.

Common mistakes

  • Building the engine first when the unknown is willingness to pay.
  • Over-advertising a concierge. Invited audience, not a press cycle.
  • Wizard of Oz on safety-critical behavior without disclosure.

Check yourself

The unknown is "will they pay $200/mo?" Fastest honest test? Concierge or a paid waitlist. Charge before you build the engine.

The unknown is whether an AI summary feels useful. Build the model first? Not yet. Wizard of Oz the experience, then decide what to automate.

Key takeaways

  • Choose concierge, Wizard of Oz, single-feature, or piecemeal — cheapest honest signal.
  • Match the method to what you must observe: pay, UX, habit, or feasibility.
  • Do unscalable manual work early. Automating first scales the wrong thing.

Further learning

Do Things that Don't Scale — Paul GrahamDo Things that Don't Scalepaulgraham.com

Map: MVPs and scoping. Next: appetite and shaping.

Riskiest assumption

An MVP tests the belief that, if false, kills the idea — written so it can fail.

MVPs and scoping

Building the smallest thing that tests the riskiest assumption, and iterating.

Pivot, persevere, or kill

Pre-declare the signals that mean continue, change the job, or stop — a clean kill is a successful MVP.