Prototypes, in three questions

// Field Guide 04

Prototypes, in three questions

Somebody on your team is going to open a prototype in a meeting this week. Maybe it's a storyboard. Maybe it's a click-through with fake data. Maybe an agent built it overnight and it runs. Same word for all three, and the room will judge them the same way.

Two Apple designers sorted this out in 1997, and most people never got past the title of the paper. Stephanie Houde and Charles Hill's rule is simple: ignore what the prototype is made of. Ask what question it answers. There are only three.

Every prototype is a question with a body. This guide sorts them by the question. Houde & Hill's own definition: "any representation of a design idea, regardless of medium. This includes a preexisting object when used to answer a design question." Their first line: "Is a brick a prototype? The answer depends on how it is used."

Answer three questions, find what you should build next ↓

The reference is here. The argument is in the memo: The Economy of Prototypes.

Drawn askew on purpose. The paper does the same, because "no one dimension is inherently more important than any other." A prototype can sit anywhere inside; position is "relative and subjective." Tap a corner to open it.

What it's made of

StoryboardA sequence of sketches showing someone using the thing in context. The paper's E4: a paper storyboard of a student taking notes on a portable device. Nobody can click it, which is the point. Concept videoA film of the product working, with nothing working. Knowledge Navigator (1987, six weeks, nothing real) and Dropbox (2007) are the canon. Landing pageA description, a price, and a button. Measures whether anyone clicks. A waitlist measures curiosity; a refundable deposit measures slightly more; a charge measures demand. ConciergeDeliver the service by hand, to a handful of people, before any system exists. The earliest Wizard of Oz: the wizard does the whole job, not just the voice. BodystormingAct the scenario out physically, with props, in the real place. The architect's computer (E10) was a weighted pizza box carried around for days.

Hover a chip for the definition.

In the wild

Prototype	What it was	What it answered
Apple Knowledge Navigator, 1987	A five-minute film of a professor and a bow-tied software agent. Made in six weeks for a John Sculley keynote. Nothing in it worked.Houde & Hill E6; Dubberly	Whether a broad audience could see the role of a machine that did not exist yet. They could. Legend later held George Lucas directed it and a Cray rendered it.
Dropbox demo video, 2007	Two three-minute screencasts of a sync product the public could not yet install, posted to Hacker News, then Digg.First Round Review	Whether people wanted invisible sync. "Within a day, video #2 was at the top of Digg, spiking Houston's wait-list from 5,000 to 75,000."
Tesla Cybertruck reveal, Nov 2019	A truck years from production, on a stage, plus a $100 refundable reservation button. The armored glass cracked live.TechCrunch	Whether there was demand for a truck shaped like that. 146,000 reservations in two days, 200,000 by day four. Mind the instrument, though: a refundable $100 measures interest. A purchase is a different number.
Architect's computer (paper E10)	A weighted pizza box, carried like the proposed device.Houde & Hill E10	Field use "revealed rectilinear form awkward," which "led to soft shoulder-bag concept." The brick argument, made literal.

Spot it

Nobody can click on it. What you have is a story (a video, a storyboard, a page) or a service with a human doing the machine's job. The feedback is about wanting, fitting, and context. If the conversation turns to buttons, you built the wrong thing or showed it to the wrong room.

Works when

The open question is demand or fit, not feasibility or form. Early, before anyone is attached. Any time being wrong about the why costs more than being wrong about the how.

Next corner

Once the role survives real people, the live question becomes Look and Feel (what is using it like?) or Implementation (can we actually do it?). The paper's space-planning team ran both next, in parallel.

Cheapest version

A storyboard and five conversations. If you need more than a day to make a role prototype, you are answering a different question.

// The honest limit

A role prototype proves people want the thing. It says nothing about whether you can deliver it, and the more persuasive it is, the more the room treats it as a plan.

Humane's 2023 launch film was the Knowledge Navigator for the AI era: an ambient companion pinned to a lapel. It raised more than $230M on that role. The shipped Pin ($699, later $499) saw returns outpace sales by summer 2024, and the assets went to HP for $116M in February 2025; the pins stopped working on the 28th (TechCrunch). The role was believed. The other two corners were never answered.

The paper called it in 1997: "Prototypes are not self-explanatory: looks can be deceiving." Role prototypes get believed. That is the point, and the problem.

Companion: The Economy of Prototypes. "A prototype with no siblings is not evidence."

What it's made of

Wooden blockA physical model at the right size and weight, with no function. Jeff Hawkins carried one in his shirt pocket for months before the PalmPilot existed. Click-throughScreens wired together so taps move between them. Real flow, fake data, no backend. Figma's default output, and now the default output of prompt-to-app tools too. Wizard of OzA person behind the curtain produces the machine's responses live. The paper's E9 hid a walkie-talkie and an RC car in a toy ball. Tesla hid teleoperators behind Optimus. Appearance modelA non-working object that looks exactly like the product, for photographs and hands. The paper's E7 used one to run market research on a communicator before it existed. Generated appLovable, v0, Figma Make, Bolt: a running front-end with seeded data, in minutes. A look-and-feel prototype that happens to compile. It has not answered the other two questions just because it runs.

Hover a chip for the definition.

In the wild

Prototype	What it was	What it answered
PalmPilot wooden model, ca. 1995	Hawkins cut a block of wood to shirt-pocket size and used "a chopstick for a stylus. He took pretend notes in meetings, and counted the steps it took to perform common tasks."Computer History Museum	Size, weight, taps per task. Zero implementation. The most famous look-and-feel prototype ever made cost nothing.
Tesla Optimus at "We, Robot," Oct 2024	Humanoid robots walked, danced, poured drinks, and chatted with guests. Bloomberg, The Verge, and Morgan Stanley reported they were teleoperated; one bartender bot acknowledged on video it was "being assisted by a human."TechCrunch	What it is like to stand next to one, talk to it, be served by it. A textbook Wizard of Oz, and a useful one. The problem was the framing: "it'll be able to do anything you want." The prototype answered the feel question. The pitch claimed all three.
GloBall learning toy (paper E9)	A walkie-talkie and a radio-controlled car hidden inside a ball, a designer behind the curtain doing speech and movement.Houde & Hill E9	The 1990s Optimus: a child's experience of a talking, rolling toy, with none of the technology.
Click-throughs, then generated apps	Clickable screens; increasingly, generated screens with running front-ends and seeded data.NN/g, 2016	Flow, hierarchy, feel. NN/g's line now applies to everything: hi-fi prototypes "often look like 'live' software to users."

Spot it

You can touch it, tap it, or hold it, and it responds, but the response is canned or a human is producing it. The data is fake. Look behind it and there is nothing there, or there is a person. The feedback is about flow, confusion, delight, size, and weight.

Works when

The role is settled enough that the open question is the experience, and the experience is the risk: hardware form factors, novel interactions, anything children or first-time users will touch.

Next corner

To Implementation when the experience depends on something unproven (latency, rendering, a model's accuracy). To Integration only when all three corners have been answered separately.

Cheapest version

Paper screens and a colleague playing the computer. The paper's first principle: "Efficient prototypes produce answers to their designers' most important questions in the least amount of time."

// The honest limit

Polish makes people believe. It teaches you very little. Sauro's synthesis of eleven fidelity studies: rough prototypes surface most of the major problems a polished build surfaces, while participants consistently prefer the polished version (MeasuringU). NN/g says it plainly: with low-fi, "stakeholders recognize work remains unfinished" and designers are "less emotionally invested." With hi-fi, the opposite on both counts. The paper, 1997: "inexperienced audiences tend to believe them to be more functional than they are just by virtue of being shown on a computer screen."

The contemporary failure is the Wizard of Oz that never leaves the building. Builder.ai marketed AI-assembled software, raised over $500M from backers including Microsoft and Qatar's sovereign fund, and was valued near $1B. The Wall Street Journal reported in August 2019 that the company "used human engineers rather than AI for most of its coding work." It filed for insolvency on May 20, 2025 (The Register). A person behind the curtain is a fine way to prototype. Six years later, with $500M raised, the curtain is the company.

Companion: The Economy of Prototypes, the section called "What a working thing does to a room."

What it's made of

SpikeA short, throwaway build that tests one technical unknown. Ugly by design. If it has a nice UI, effort went to the wrong corner. BenchmarkMeasure the hard number: frames per second, latency, error rate, tokens per dollar. The paper's E3 measured how complex a 3D scene a 1990 Mac could draw, which set the ceiling of the product. DemonstratorA one-off working article that proves the physics or the architecture, with no intention of being the product. Boom's XB-1 carried zero passengers and flew supersonic. Cost modelThe spreadsheet version of "can we build it": unit cost, margin, supply. Not in the 1997 paper; added here because "can we deliver this" includes the arithmetic. See the pricing guide. Technical specWritten architecture that lets engineers find the impossible part before anyone builds it. Cheapest implementation prototype there is, and the most skipped.

Hover a chip for the definition.

In the wild

Prototype	What it was	What it answered
Boom XB-1, Jan 28 2025	A one-off, human-piloted demonstrator jet. Eleven flights from March 2024; Mach 1.122 at 35,290 ft over Mojave.CompositesWorld	Whether a small private company can build a supersonic airframe at all, and whether Overture's bets (carbon composites, digital stability augmentation, an augmented-reality landing vision system) work in the air. Nobody will ever ride in XB-1. Scholl: "A small band of talented and dedicated engineers has accomplished what previously took governments and billions of dollars."
Figma, 2012 to 2015	Evan Wallace's custom WebGL renderer, years before a product shipped; revenue did not start until 2017.TechCrunch	Whether a professional design tool could run in a browser tab at native speed. Figma's CTO on first seeing it: "wow, this is actually legitimate, this is as fast as a native application."
3D space planner, rendering test (paper E3)	Working code on a Macintosh IIfx.Houde & Hill E3	How complex a scene the hardware could draw. The answer set the ceiling of the product before anyone designed a screen.
Digital movie editor (paper E11)	A C program for marking movies and attaching scripts.Houde & Hill E11	Implementation. But the audience "became focused on problems with the user interface, such as the multiple cascading menus... In these cases, discussion was less productive." The canonical wrong-corner reaction.

Spot it

It works and it's ugly, or it works and nobody outside engineering has seen it. The output is a number (frames per second, latency, cost per unit, error rate) or a yes/no. A nice UI here means somebody spent a day on the wrong corner.

Works when

The idea dies if one specific technical or economic thing is not true, and nobody knows yet whether it's true. Also the corner to return to whenever a look-and-feel prototype faked something hard: a model, a sensor, a margin.

Next corner

To Look and Feel once feasible, because "feasible" and "usable" are different findings. To Integration when all three have been answered.

Cheapest version

A one-page spec and an afternoon spike on the single scariest assumption. Measure it. Write the number down.

// The honest limit

An implementation prototype proves the thing can be built. Whether anyone wants it, or can use it, is still open. It also has a failure mode the other corners don't: being judged by the wrong audience on the wrong dimension, like the movie editor's cascading menus.

The 2026 version is AI-shaped. Generated code arrives looking finished, and the constraint has moved from the build to the review. LinearB's 8.1-million-PR dataset: AI-generated changes merged within 30 days 32.7% of the time against 84.5% for human-authored, and waited 16+ hours for a reviewer against about 200 minutes (LinearB, May 2026; method unpublished). The implementation question used to be "can we build it." It is increasingly "can anyone tell whether what got built is right."

Cross-links: Pricing, in five moves (move 1, for the cost model). The Economy of Prototypes for the review queue.

You earn your way to the center. The paper, on why its space-planning team built three separate prototypes instead of one: "It was more efficient to wait on the results of independent investigations in the key areas of role, look and feel and implementation than to try to build a monolithic prototype."

What it's made of

Working systemReal code, real data, real look, real use, one artifact. The paper's E13 sound browser and E14 pile metaphor. Expensive, usually late, and the only way to test the seams between the three corners. Field pilotThe working system in real hands for a bounded period. Past this point you are testing operations, not design; see "Not prototypes" below. Generated app (warning)Running code plus seeded data reads as integration. Most of the time it has answered one question while looking like it answered three.

Hover a chip for the definition.

In the wild

Prototype	What it was	What it answered
Sound browser (paper E13)	A working C program for browsing audio with visual markers; visual design iterated in a pixel-painting tool alongside the running code.Houde & Hill E13	All three dimensions, with the team able to change the look while the thing ran. The paper's model integration.
Garment history browser (paper E15)	A SuperCard working system for fashion designers. Two months to build.Houde & Hill E15	The complete experience, narrowly. The designers "had to invest a large amount of time in making the prototype, yet had only been able to support a very narrow role compared to the breadth shown in the animation" (E8, a 20-minute look-and-feel animation that preceded it).
Amazon Just Walk Out, 2018 to 2024	Cashierless stores: cameras, shelf sensors, computer vision, real shoppers, real money, six years in production.Axios	Then, April 2024, Amazon pulled it from its Fresh grocery stores in favor of Dash Carts. Reports put over 1,000 workers in India reviewing video; Amazon said the team was "way less than 1,000," worked "after the fact" to train the models, and verified receipts in "a small percentage of cases." Either way: shipped as integrated, with a person still partly behind the curtain.

Spot it

Real data, real code, real look, real use, all in one artifact. Expensive. Usually late. If it exists in week two, either the team skipped the corners or you are looking at a click-through that happens to run.

Works when

Each corner has been answered separately and the remaining risk is in the seams: the tradeoffs between what people want, what feels right, and what can be built. The paper: it helps "balance and resolve constraints arising in different design dimensions to verify that the design is complete and coherent."

Next corner

There isn't one. Integration is the last stop before the product. The only moves from here are ship, or go back to a corner.

Cheapest version

There is no cheap version. That is the tell. If it was cheap, check which corner it actually answered.

// The honest limit

Integration is the most expensive prototype to build and the easiest to fake. Two months of SuperCard bought a narrower role than a 20-minute animation. Just Walk Out ran six years as a shipped product and still had humans checking the machine; when Amazon retreated, it retreated to a cart with a screen, a smaller integration that held.

And the fake is now free. Lovable reports a million new projects a week; Vercel's CEO says 30% of apps on the platform came from agents (TechCrunch, TechCrunch; both company-reported). A generated app with real-looking data, a real-looking UI, and running code reads as integration. Most of the time it has answered one question while looking like it answered three. The paper's fourth principle is the defense: "Be clear about what design questions are being explored with a given prototype, and what are not."

Companion: The Economy of Prototypes. The integrated-looking demo is "an argument that learned to run."

// One project, three prototypes

The paper's own example, on one screen

An Apple team set out to build a 3D space planner: virtual furniture in a virtual room. Instead of building one prototype of the product, they built three small ones at the same time, one per corner. None of them looked like a product. None of them looked like each other.

// Role

A slide show.

Fake screens of someone furnishing a room, walked around the company to see if anyone wanted the thing to exist.

It answered

Is this worth building?

// Look and Feel

A click-through with pre-drawn chairs.

You could grab a chair by its "handle box" and spin it. Nothing behind the screens.

It answered

Does moving furniture this way feel right?

// Implementation

An ugly rendering benchmark.

Working code pushing a 1990 Mac as hard as it would go, watched by nobody but engineers.

It answered

How much room can the hardware draw?

E1 to E3 in the paper. Three prototypes, three audiences, three answers, and a product decision made without a monolith.

The same team today could generate the whole thing on day one. It would demo beautifully. It would have answered one question out of three.

// Not prototypes

Words people use as synonyms that sit outside the triangle

MVP

A Minimum Viable Product is a product, shipped to paying or committed users, optimized for learning. Many famous "MVPs" were role prototypes (Dropbox's video, the Cybertruck reservation page): nothing viable was sold. Rule of thumb: if money or commitment changes hands, it's an MVP; if a question gets answered, it's a prototype; the same artifact can be both.

Demo

A demo is a prototype with a different job: it exists to win a decision, and any question it answers is a side effect. The paper's third principle is the tell: "The necessary resolution and fidelity of a prototype may depend most on the nature of its audience." A demo's audience is the approver. The companion memo is about what happens when demos get free.

Pilot

A pilot is the real product in a bounded deployment. It answers an operations question (does this work at scale, with real users, inside real constraints), which is past the triangle. Deloitte's finding that only 25% of organizations have moved 40% or more of their AI pilots into production (n=3,235, Jan 2026) is a pilot problem, not a prototype problem, though it is often caused by skipping the Role corner.

Proof of concept

Usually an implementation prototype with a more formal name. Keep it in its corner; don't let it pretend to be integration.

// What are you actually prototyping?

What should you build next?

Three questions. You get the corner, the cheapest way to test it, and the one thing that prototype will never tell you.

1. If this idea fails, what's the most likely reason?

2. Who most needs to be convinced next?

3. What have you already proven?

Cheapest tool

What it will not tell you

If you were about to build the whole thing, build the corner instead, then build the other two. Three small prototypes beat one monolith (Houde & Hill, E1 to E3).

That is the reference. The argument, what free demos do to a room, is in the memo: The Economy of Prototypes.

Prototypes, in three questions

Prototypes, in three questions

What does this do in someone's life, and do they want it there?

What it's made of

In the wild

What is it like to use?

What it's made of

In the wild

Can it actually be built, and what will it cost?

What it's made of

In the wild

All three at once: is the whole thing coherent?

What it's made of

In the wild

The paper's own example, on one screen

Words people use as synonyms that sit outside the triangle

What should you build next?