Earn the first prompt

Nobody churns because your model is 4% worse than the frontier. They churn because their first prompt was “tell me a joke,” the answer was fine, and nothing about the experience suggested there was more. The blank box outsources your entire value proposition to the user’s imagination, and most users — reasonably — don’t bring imagination to a product they haven’t decided to trust yet. The first prompt is yours to design, not theirs to invent.

Prompt starters are the blunt instrument and they work. ChatGPT and Claude both greet new sessions with concrete example prompts, and the good ones are specific enough to be tried verbatim — “summarize this PDF” beats “get creative” every time. Perplexity’s home screen leans on trending and suggested questions, which does double duty: it seeds a first query and demonstrates that the product answers questions rather than chats. Vague starters (“brainstorm ideas!”) are worse than none, because they confirm the suspicion that the product doesn’t know what it’s for.

The stronger move is contextual: don’t make people come to the AI, put the AI where the problem already is. Notion AI appearing when you hit space in an empty block, Gmail offering to draft a reply on the message you’re reading, Figma surfacing AI actions on the selection — these earn the first prompt by making it nearly zero-cost and obviously relevant. The person never has to translate their situation into a query; the entry point carries the context.

The first result matters more than the first prompt. Route new users toward tasks the system is genuinely good at, even if that means hiding the impressive-but-flaky capabilities early. Linear’s approach to product craft applies here: better to do three things excellently on day one than to hand someone a menu of forty where a third are coin flips. A guided first run that walks someone to one real success beats a feature tour of ten hypothetical ones.

And retire the training wheels. Starters and suggestions are scaffolding for people who don’t yet have their own use cases; once someone has a history of real prompts, showing them generic suggestions reads as the product forgetting who they are. Earning the first prompt is a cold-start pattern — the whole point is to make itself unnecessary.

The strongest case against all this scaffolding is that it patronizes the exact users who matter most. Power users arrive with intent, and every starter chip between them and the composer is friction; Google won the search wars with a blank box and nothing else. There’s a subtler cost too: curated starters narrow the perceived surface of the product — people conclude the AI does the six things on the cards and nothing more — and contextual entry points sprinkled through an interface can curdle into the “AI confetti” that users now actively resent. Shipping forty suggestion surfaces is also real engineering spend that a leaner competitor puts into model quality. All fair. But the blank box only worked for Google because everyone already knew what search was; nobody yet shares a mental model of what your AI is for. Scaffolding that retires itself costs power users one glance and saves everyone else the product. The failure isn’t the starter — it’s the starter that never leaves.

Measure it at the first session and the fourth. The headline metric is first-result quality: what fraction of new users’ first prompts land on tasks the system handles well — track it by classifying first prompts against your known-good task list, and watch the trivial-prompt share (“tell me a joke” equivalents) shrink as your starters and entry points improve. Good looks like a majority of first sessions ending in an output the user acts on: copies, saves, sends, or follows up substantively rather than abandoning. Then measure the handoff: by session four or five, organic prompts should dominate and starter-clicks should be a rounding error — if veterans are still clicking suggestions, your product has no depth to graduate into; if new users never click them, your starters are decoration. The single sharpest number is second-session return rate segmented by first-result quality. When a good first result predicts return and you’re moving the share of good first results, the principle is paying rent.