Capability before the cursor

Tags: onboarding, affordances, chat

A cursor's only affordance is type anything, which is the same as declining to say what the product does. So a new user guesses — and guesses wrong in both directions.

It is a calibration problem, not an inspiration problem

The blank prompt is usually framed as a cold-start difficulty: people do not know what to write, so give them something to write. That framing is too small, and it produces solutions that address the symptom.

The actual failure is that a user cannot form a correct model of the product's range from a text box, so they settle on a wrong one:

• Under-asking. They try something trivial, it works, they conclude this is a toy with a chat skin, and they do not come back. The product was never shown.
• Over-asking. They try something beyond it, get refused or get something wrong, and conclude it is broken. Same outcome, opposite cause.

Both are calibration failures. Both belong to the interface, because the user had nothing to calibrate against.

Which is why a row of example prompts only half-helps. It gets somebody started, and it communicates nothing about where the edges are — so the over-asker is unaffected, and the under-asker is now under-asking with more confidence.

Four, not twenty

The number is the decision, and it is the part that transfers.

Twenty suggestions is a list to work through. Four is a shape to grasp. A person reads four in two seconds and comes away with a claim about the product: this is for these kinds of things. That claim is the thing you were trying to deliver. Twenty items deliver no claim at all — they deliver a scrolling problem, and the user reads the first two and forms the same under-calibrated model they would have formed from nothing.

So the work is not collecting examples. It is choosing the small number that describe the territory, which means each one has to be doing different work: different kind of task, different length of output, different level of ambition. Four that are variations on each other are one suggestion with three wasted slots.

Show the edges

Half the model is what the thing will not do, and almost nothing ships it.

This does not require a list of limitations — a wall of caveats is its own failure. It requires that the displayed range have a visible boundary: the most ambitious example is genuinely near the limit, so a user extrapolating from it lands close to reality rather than far past it.

The inverse is the common mistake. Suggestions chosen to be impressive set an expectation the product cannot meet on the second request, which is the over-asking failure manufactured by the onboarding that was supposed to prevent it.

When a cursor is right

A search box needs no suggestions, because everybody already knows what search does — the capability was established by thirty years of convention. A prompt box inside a spreadsheet inherits a great deal from the spreadsheet. The cursor is the right control once capability is understood; it is the wrong thing to meet somebody with.

And for a returning user it is usually right again, which argues for this being a first-run and empty-state concern rather than a permanent fixture. Suggestions that never go away are a product that never stops introducing itself.

Grounded in

Poppy opens on four one-tap actions: plan a lesson, write to parents, build a quiz, adapt content to a reading level.

Underneath it is structurally a chat app — streaming responses, persisted threads, the whole shape. The part that made it usable is the part that is not chat. A teacher with ninety seconds between classes and a blank prompt has a staring contest, not a conversation; the same teacher shown four actions knows within two seconds whether this product is for them today.

Four was a claim about the shape of the work, chosen from a much longer list of things it could do. The longer list would have been more accurate and would have communicated less.

Anti-patterns

• A blank prompt as the first-run state. The default, and a decision not to say what the product does.
• Twenty suggestions. A list instead of a shape.
• Suggestions that are variations on one idea. Four slots spent delivering one claim.
• Only easy examples. Miscalibrates downward; the user concludes it is a toy.
• Only impressive examples. Miscalibrates upward; the second request fails and the product looks broken.
• Nothing about the edges. Half the model, routinely omitted.
• Rotating suggestions. A changing set prevents a stable model from forming, which is the one thing this is for.
• Suggestions that never leave. A returning user with a formed intent is being interrupted by the onboarding.

The smallest version worth building

Four actions on the empty state. Each a different kind of task. The most ambitious one genuinely near the limit of what the product does well.

Then remove them once the user has a history, because by then the capability is established and the cursor is the better control.

Related patterns

• Wizard or conversation — the next question after this one: once they have picked a capability, whether the interaction should be a conversation at all.
• Show the machine listening — the same obligation for continuous input, where the interface also has to say what is happening rather than assume it is obvious.
• Design the failure state first — over-asking produces a refusal, and the refusal screen is where a miscalibrated user is either recovered or lost.