Introduce capabilities at the right moment

What it is

Capabilities will not be groped out of the dark unaided. The system has to deliver a neighbouring capability at the moment it is usable. A course-registration voice that has just added a class, then says “if this one fills, you can say ‘join the waitlist’” — waitlist sits on the same goal as the last act, so it can land. Dumping waitlist, transcripts, fees, and housing into the opening is a catalogue delivered before anyone has a goal. Wrong moment, and the introduction might as well not have been said.

Why it happens

A new capability enters the plan only when it can hang off a goal that is already live. Right after “add course,” waitlist is a branch on the same road; working memory still holds the course name and the afterglow of success, so a short intro can attach. Before any goal exists (just woken), the same sentence has no hook and is filtered as opening noise. After the goal has moved (the person is asking when the cafeteria closes), waitlist is stealing the current task.

So introduction is not better for being earlier or more complete. It has to be relevant, and needed but not yet used. Too early: no hook. Too late: the person has already hit a wall on the wrong path, or has left. Proactive means the system speaks first, not that it waits for “what else can you do.” Proactive on the wrong task is an interruption.

Studying it

Hint-timing experiments: the same neighbouring capability, inserted in four places — opening, after a related success, after an unrelated act, after a miss. Dependent measures: whether the capability is used next, barge-in on the intro sentence, interference with the current task (extra turns, abandonment). A further independent variable is whether the intro shares an entity with the current slot (this course versus another course).

In logs, align “intro played” with “first use of that skill” and the gap between them. A short gap inside a similar task means timing was right; many intros and first use still coming from a completely different entry means the intro is spinning. Do not tell lab participants “you will be tested on the feature you just heard” — that turns an incidental listen into deliberate encoding.

Where it stops holding

Safety or compliance capabilities that must be known before the act (“before you submit you can say ‘I want to withdraw’”) cannot wait until after success; consequence pins the moment earlier. When the user is explicitly exploring (“I just got this device”), a very short map at the open is reasonable: the task is orientation. Repeating an intro for a capability already used becomes a tax. In public, one more intro sentence is one more exposure; timing also has a social cost.

Applying it

• Tag every skill you want discovered with a trigger situation: which goal just succeeded or was just mentioned licenses this sentence. Skills with no trigger do not go in the opening.
• Hang the intro on the current entity (this course, this bill). Do not advertise a second skill in the same breath.
• Cap proactive intros of the same skill per account; stop after use or skip.
• How to check: usage of the skill within N turns of the intro, and whether the current task got longer. Usage flat and the current task longer means the hook or the moment is wrong; move the intro to a success turn that actually shares an entity and measure again.

Related

• Same group: M2.06.1 A voice interface has no scannable feature list · M2.06.3 A screen is the most effective supplement · M2.06.4 People ask for help after failure · M2.06.5 Capability intros need speakable examples · M2.06.6 “What else can you do” is a bad question to answer
• Nearby: M1.12 Opening and closing a dialogue · M2.01 Information density of prompts · M1.05 Topic switching and task nesting
• Search terms: just-in-time capability introduction · situated hint · feature discovery