AI Prompt Box Design: Why Ask Anything Fails

Good AI prompt input design tells a new user what to type before they type anything. A single empty box with the placeholder Ask anything does the opposite: it hands the hardest part of your product, knowing what a good request looks like, to the one person who has never used it.

For an AI founder this is not a styling question. Your output is only as good as the input, and right now you are probably the only person who knows what good input looks like. That is why the product lands in your demos and falls flat when a stranger signs up alone.

A research tool whose first screen is one box

Picture an AI research tool for product marketers. You sign up, confirm your email, and land on a white page. In the middle sits one text field, about 600 pixels wide, with grey placeholder text that reads Ask anything. There is a paperclip icon on the left and an arrow button on the right. Nothing else is on the screen.

A new user types what they would type into Google: what are competitors saying about our pricing. Eight words. The answer comes back in five seconds. It is three tidy paragraphs about value-based pricing and the importance of monitoring the market. Nothing in it is wrong, and nothing in it is about their competitors. They try again with a shorter question and get the same careful tone. On the third try they close the tab. Their conclusion is simple: this is another wrapper.

Now picture the founder at the same box. He pastes in about 200 words. He tells the model to act as a pricing analyst. He names four competitors and says to use only their public pricing pages and changelogs from the last six months. He asks for a table with plan name, price per seat and what changed. He adds one example row. What comes back is the thing the landing page promised. Same product, same box, same model. The difference is that he knew how to ask.

What Ask anything tells a stranger before they type

An empty box is a message, and the message is: you figure it out. It says there is no preferred use, no right length and no sense of what this product is best at. A search bar trains people to type short. A box that looks like a search bar, shaped like one and sized like one, gets short input. Short input gets generic output.

Founders often defend the empty box as minimalism. It is not. Minimalism removes what the user does not need. The empty box removes what the user needs most, which is a picture of a good request. Jakob Nielsen has called this the articulation barrier and judged that about half the population in rich countries is not articulate enough to get good results from current AI bots. Your buyers may write well. They still do not know your product's vocabulary.

The box also hides your range. If the tool can compare pricing pages, summarise sales calls and draft a battlecard, a new user learns none of that from a blinking cursor. They learn only what they already guessed, and they guessed search.

Why the demo closes and the signup leaves

Watch what happens in the founder's week. He runs a dozen demos and most of them close, because he is the one typing. Self-serve signups sit at a few percent. He reads that as proof the product is great once people see it properly, so he books more calls and starts drafting a job post for a salesperson.

That reading is expensive. The salesperson will not know how to drive the box either, not after a two-week ramp. Meanwhile every signup who types eight words and leaves still costs you inference on three flat answers. Measured on AI companies, product builders average around 52 percent gross margin against 70 to 80 percent for traditional software, so a wasted session hurts more here than it would in a classic SaaS tool.

Measured on SaaS and product-led companies, activation runs 20 to 40 percent for most products, and a ten point improvement typically drives a 15 to 25 percent increase in free-to-paid conversion. For a prompt-first product, the first request is the activation moment. If the first answer is flat, there is no second session to improve.

The prompting guide in the docs does not fix it

The usual fix is a page called How to write great prompts, linked from the help menu. It lists tips like be specific, give context and set a format. This is the part most advice gets wrong. A guide asks the user to leave the product, read a lesson and come back better at a skill they did not sign up to learn. Almost nobody does that on day one, and the ones who do are the users who were going to be fine anyway.

The guide also puts the blame in the wrong place. If people need a tutorial to get a good answer, the tutorial is describing a missing part of the interface. The fix belongs on the screen where the typing happens, at the moment of typing, in words about this product rather than about prompting in general.

Placeholder text that does a job

The placeholder is the cheapest change you can ship. It should show one real request, written the way your best user writes. Compare Ask anything with: Compare pricing changes for Acme, Brightly and Northwind over the last six months, as a table. The second one teaches the length, the kind of task and the output in one line. People copy the shape of what they see.

Three rules keep it honest. Make it a full request, not a hint like Try asking about competitors. Make it one your product answers well every time, because some users will run it exactly as written. And rotate it between two or three of your strongest jobs so returning users see the range.

Scaffolds beside the box, not inside a manual

A placeholder disappears the moment someone types. The parts of a good request need to stay visible. That is the job of scaffolds: small controls around the box that carry the structure your 200-word prompt carries in prose.

The most useful scaffolds for a first run are these. Starter cards under the box, each one a finished request with real content that runs in one click. A small row of chips for the things your best prompts always include, such as Competitors, Time range and Output as table. A source picker so the user says where the answer should come from instead of hoping. And a line under the answer that shows what the product assumed, with a link to change it.

Starter cards are the strongest of these, because they are the product showing what it can do by doing it. A user who clicks one sees a remarkable answer in the first minute and learns the pattern without reading a word about prompting. The first response in a chat interface matters just as much, but it cannot rescue a request that was never going to work.

One box, before and after, to ship this week

Here is the same research tool rebuilt without adding a single new model call. You can do each step with the team you have.

• Pull your own last 20 good prompts, the ones from demos. Mark the parts that repeat: the role, the sources, the time range, the format.

• Replace Ask anything with one of those prompts, shortened to a single line that still names real inputs and an output.

• Put three starter cards under the box. Each is a full request with real names and a Run button. No lorem ipsum, no Example Company.

• Turn the two parts that repeat most into chips or fields beside the box. If every good prompt names a time range, give time range its own control.

• Make the box taller than one line and let it grow. A search-bar shape asks for search-bar input.

• Under every answer, show one sentence of what the product assumed and one suggested follow-up request.

• Watch five new users try it without you on the call. Count words in their first request. If the median is under 15, the box is still teaching search.

Before, the first screen asked a stranger to invent your best use case. After, it shows them three of them and lets them run one. The founder's 200 words did not go away. They moved from his head onto the screen. If you want a fuller picture of how a first screen should hand off to the rest of the product, the notes on empty state design apply to a prompt box more than most people expect.

Where Studio Maydit fits

Studio Maydit designs the screens where AI products get used for the first time, which in a prompt-first product means the box, the few seconds after it, and the answer that follows. We start from the requests the founder types in demos and turn them into starter cards, scaffolds and defaults a stranger can use alone. Our AI product design work runs as a fixed scope over three to four weeks and ends with a diagnosis of what is still leaking. If you are weighing options, we also keep a list of studios that work on prompt and input UX. If your product only lands when you are the one typing, book a 30-minute call with Studio Maydit.