The empty prompt problem: what to show first

A cursor is not an interface

Open a new AI agent for the first time and you usually get one thing: a text field, a blinking cursor, and placeholder copy that says some version of "Ask me anything." That's the entire onboarding. No tour, no examples, no scaffolding — just an empty box and an implicit demand that you already know what this tool is good for, phrased correctly, before it will show you anything.

That demand is backwards. Jakob Nielsen's decades-old usability heuristic — recognition over recall — says people are far better at picking a familiar option from a list than dredging one up from memory. A blank prompt box does the opposite by construction: it asks a first-time user to recall what the product can do, in their own words, before they've seen the product do anything. Nielsen Norman Group's research on new gen-AI users found the fundamentals haven't changed just because the interface is conversational — first-time users of Chinese gen-AI chatbots in their study struggled to understand basic functionality, had real questions about what the tool could do, and the tools routinely failed to answer those questions at the moment they came up. NN/g's practical takeaway is almost anti-onboarding: skip the tutorial when you can, because people don't want to read instructions before they get to try something, but back that off-ramp with contextual help the bot can produce on demand, and give the system real self-knowledge of its own capabilities so it can answer "what can you do?" honestly.

The cost is measurable, not theoretical

This isn't an abstract design complaint. In a UX Collective piece on empty states in AI products, designer Adi Leviim traces the pattern across the current generation of chat products — nearly every AI tool launched since 2023 ships some variant of the same centered text field and "ask me anything" placeholder — and reports that for a Chrome extension he co-built, roughly 70% of new installs never opened a second session, with the empty state as the single biggest reason people bounced. His diagnosis matches Nielsen's: the products people came back to were the ones that showed a worked example before handing over a bare cursor. That's one product's numbers, not an industry benchmark, but it's a concrete illustration of what "recall before recognition" costs in practice — quietly, in the sessions that never happen, which is exactly the kind of failure a founder or PM won't see in a support queue.

What the current crop of products actually does about it

A few vendors have moved past the bare text field, and their approaches cluster around the same idea: replace open recall with a shortlist to recognize.

Configured starter prompts on the welcome screen. Microsoft's Copilot Studio lets a builder configure up to six suggested prompts that appear on an agent's welcome page before the user types anything — and when an agent is generated from a description, Copilot Studio can draft those suggestions automatically from the agent's own instructions. The suggestions are shortcuts into real capability, not decoration; clicking one runs the underlying prompt immediately.

Role-scoped starter prompts bundled with the tool. OpenAI's June 2026 "Codex for every role" release packages role-specific plugins — Sales, Data Analytics, Product Design, Creative Production, Investment Banking, and Public Equity Investing at launch — each bundling starter prompts, app integrations, and workflow guidance for that job function, so a new user in one of those roles gets a menu of things Codex is known to do well rather than a blank field. Details are in OpenAI's Codex role-plugin release notes.

Structured commands as a recall aid. GitHub Copilot Chat takes a different tack: instead of (or alongside) example prompts, it exposes a fixed vocabulary — slash commands like /fix and /tests, and chat variables like #selection and #codebase — that a user can discover by typing / and reading a list, per GitHub's Copilot Chat documentation. That's recognition over recall applied to the input itself: you don't need to remember the right phrasing if the product hands you a menu of verbs.

All three are the same move: turn "what do I even ask this thing" into a pick-one-of-these moment, ideally with the picks scoped to what that specific user is likely to need.

Designing the first sixty seconds

A few things follow directly from the research and the shipped examples above:

• Show, don't prompt-for. A worked example — a filled-in result, a sample output, a completed run — teaches the tool's range faster than any placeholder text can describe it. Leviim's observation that returning users saw an example before a cursor is the sharpest version of this.
• Scope suggestions to context, not the whole product. A generic "here's what I can do" list regresses toward the same recall problem at one remove. Microsoft's per-agent starter prompts and OpenAI's per-role plugins both narrow the menu to what's actually relevant to this user, which is what makes six items enough.
• Make the tool able to explain itself. NN/g's finding that users had real questions the bot couldn't answer at the right moment points at something onboarding screens can't fix alone — the agent itself needs an honest, current account of its own capabilities and limits, answerable mid-conversation, not just in a help doc written before the last feature shipped.
• Don't gate the first action behind a tour. Every one of the mechanisms above — starter prompts, slash commands, welcome-page suggestions — gets the user to a real result on the first click. None of them ask for a walkthrough first.
• Treat the empty state as a product surface, not a placeholder. It's the one screen every single user sees, and for a meaningful share of them, per Leviim's numbers, it's also the last one. That's worth the same design attention as any screen with a conversion number attached.

None of this requires a protocol or a spec — it's ordinary interface design applied to a new kind of blank page. The tools that get it right aren't doing anything exotic; they're just refusing to make the first-time user do the hardest part of the interaction — deciding what's possible — entirely from memory.

The empty prompt is one instance of a broader pattern in how agents earn a human's trust before they've done anything to earn it yet; see the tracker for how the surrounding protocols and patterns are maturing.