A prompt must say what you can do in one sentence

What it is

The system's spoken prompt is not a greeting. Before it yields the floor, it has to install the one thing that can be done now into the listener's reply plan. That thing has to fit in one sentence: one predicate, one currently executable act. “Say your clinic card number” is one sentence. “You can book, reschedule, or look up results, and also hear waiting-room instructions” may still be one grammatical sentence, but it packs two unrelated jobs; the user is handed a catalogue, not an instruction. The claim is about system speech as instruction density, not about how many options a person can remember. Once a prompt starts enumerating a list, the load has already slipped from “get the instruction” to “hold a menu.”

Why it happens

In speech, a usable instruction maps onto one intonational phrase. At the phrase boundary the listener forms a schema for “what I answer with,” then speaks. If the prompt installs a single schema before handing over the floor — give a card number, pick morning or afternoon, name a destination — planning cost for the next turn is near zero. Two predicates in one sentence force a task choice and then a wording choice; speech is delayed, and the wrong schema is often the one that ships.

This is the Gricean maxim of Quantity on the system side: supply just enough for the user's next act. Extra propositions are not heard as courtesy; they are heard as “the question is not over.” A kiosk that says “say your name and appointment slot” still has one action schema: two slots of the same form. Adding “you can also ask for department locations” installs a second schema that crowds the first out of the plan. The unit is number of predicates, not number of periods.

Studying it

Immediate probe after the prompt: play the prompt, stop, ask “what can you do now / what were you about to say.” Score whether the intended act appears in the answer, not word error. Independent variables: number of predicates (one act, two slots of one act, two unrelated acts), and whether the target act sits first or last. Dependent measures: probe hits, latency to speech, whether the spoken intent matches the designed intent.

A control writes the same capability as “please say X” versus “this system can help you with a number of services, including….” If the lab lets people glance back at a transcript on screen, the measure is reading, not auditory instruction. Field work can interrupt at a check-in kiosk or a parking-payment pillar the instant the prompt ends, to block post-hoc rationalization.

Where it stops holding

Legally required notices (privacy, fees) cannot be crushed into the action sentence, but they should be spoken separately and skippable. A necessary constraint on the same act (“weekdays only”) still belongs to that schema, not a second task. When the user already walks in with a plan (“move it to Wednesday”), the prompt can shrink to an acknowledgement; the one-sentence rule is for system-first openings, before the user has an action plan. Splitting three tightly bound slots into three “please say…” turns shreds a schema that one sentence could have held.

Applying it

• For every state where the system speaks first, write one action sentence: one verb, pointing at what the user must hand over next. Count predicates; more than one means split the state or change the question.
• Move policy, brand greeting, and capability catalogues out of the action sentence. Required notices get their own sentence and a skip.
• “Name and appointment slot” may share a sentence; they fill one form. “Book a visit” and “how do I find the department” may not.
• How to check: five-person probe per prompt — after hearing it, “what do you say now.” Rewrite until the target act is in the answer. Listen once more to confirm a second verb is not stealing the schema.

Related

• Same group: M2.01.2 Overlong prompts get cut off or forgotten · M2.01.3 Novices and experts need different prompt lengths
• Nearby: M1.12 Opening and closing a dialogue · M2.06 Discoverability and help · M1.02 Memory load of screenless interaction
• Search terms: one-sentence prompt affordance · spoken prompt · system initiative