Examples draw the range of what is possible, so diversity matters more than count

What it is

An example is not a compressed manual. It is a sample from the capability space. Five examples that all teach “write an email” mark a point. Three that land on write, revise, look-up, and change-of-format mark a region. What people learn from examples is “things like this belong here,” so diversity outruns count. Ten homogeneous chips do not enlarge the map; they stamp the same plot of land harder.

This is not about whether examples ease the first sentence. It is about examples as boundary markers: the sampling policy decides how large the user’s map becomes.

Why it happens

People induce a category from a handful of members. The more alike those members look, the narrower the inferred category. That is a stimulus-sampling problem: if every stimulus in an experiment comes from one subclass, subjects lock the conclusion onto that subclass. Product chips are the stimuli.

Diversity has to be sampled on the dimensions people actually split when they decide, not on surface wording. “Write a weekly report,” “write an email,” and “write a notice” still count as one point — short prose. “Turn a table into a chart,” “turn a meeting recording into todos,” and “point at a region of an image to change” split input modality, output form, and object of action. Counting near-paraphrases creates the illusion of “we shipped many examples.”

Studying it

Have designers produce a capability grid (modality, object, verb, output form). Code every on-screen example into that grid and count occupied cells, not example rows. Then compare between groups: high-coverage sampling versus homogeneous repeats, and watch the entropy of categories people spontaneously try in the next 15 minutes. Independent variables: occupied-cell count, whether the set includes an out-of-scope “cannot do this” example. Dependent variables: number of categories attempted, rate of out-of-scope tries, whether success concentrates in cells that were exemplified.

Coding must be done by someone who did not write the examples, or designers will treat paraphrases as different cells. On the log side, topic-model or hand-code the first N user requests and see which cells they land in — that is the range the examples actually drew.

Where it stops holding

A tool whose job is one narrow act (summarise one kind of contract) does not need cross-class diversity; homogeneous examples are an honest range statement. Once experts already know the map, the marginal return on diversity drops; they want an editable skeleton, not a new point. Out-of-scope examples sharpen the boundary, but if the product strategy is to encourage exploration, early negative examples can drive people off paths that are in fact still open.

Applying it

• Author examples as a coverage problem: keep a capability grid, and each shipped example must occupy a still-empty cell. Ban padding the count by rephrasing.
• Include at least one example in a combination users may not invent but the product does support (ask questions of this image, not another “make it shorter”).
• Periodically rewrite the grid from real first-utterance distributions. Cells users discovered on their own that chips never showed mean the sample was biased, not that the chips can be deleted.
• Check: drop the count metric; report occupied cells over planned cells. A release that adds rows without adding cells is a regression. Then sample 50 first utterances from new users and count how many land outside exemplified cells — a very low share means the range has been drawn shut.

Related

• Same group: L2.09.1 Natural-language interfaces have nothing to scan, so discoverability is worse than in graphical ones · L2.09.3 People treat seen examples as a ceiling; a narrow set shrinks real use · L2.09.4 Suggestions that appear only after failure arrive past the point of giving up · L2.09.5 Capability hints must move with the dialogue; a one-shot start-screen hint does not cover later turns
• Nearby: L2.03 Examples and Template Guidance · L2.02 Discoverability of What Can Be Said · L2.01 Openness of Natural-Language Commands and Its Cost
• Search terms: example diversity over count · stimulus sampling · capability space coverage