Templates turn open input back into structured blanks, at the cost of openness

What it is

The product announces a natural-language entry — operate by speaking — then folds frequent tasks into slotted templates. The interaction grammar moves from “any sentence” back to “fill the cells.” Open-to-form conversion names that recapture: the template is not a help layer on an open channel, it re-closes the channel into a form. What you gain is checkability and a start. What you spend is the slice of expressive freedom natural language was bought for — intents with no pre-built cell have nowhere to go.

A slot roster shaving expression is people still in the language channel, saying fewer dimensions. Grammar recapture is more thorough: they are no longer in the language channel. They are filling a form.

Why it happens

The value of open input is delayed binding — structure appears at speech time, so rare intents, compound intents, and mid-flight changes of mind still have somewhere to sit. A template pulls binding forward to design time: which cells exist is decided by the product first. That is ordinary GUI practice. A generate feature wearing templates rebuilds a form outside the model, and the model’s combinatorial power is discretized again at the entry.

Products often sway between the two without declaring it: the ad says “ask anything,” the main path is a template gallery. People prepare an open narrative for the ad, then are asked to tear it into cells. The conversion itself costs — the narrative is shredded, relations (“compare price first, then lead time only for the top three”) will not fit, and get dropped or stuffed into Other.

Studying it

Give the same set of compound tasks (including cross-slot relations like “first… then only for…”). Three groups: pure open, pure template gallery, open and template side by side with a forced choice of path. Code whether artifacts keep the cross-slot relation, whether people wanted to change grammar mid-way, reasons for the path chosen. Independent variables: how well the template covers the task, whether conversion is reversible (can filled values dump back into one natural sentence). Dependent variables: relation preservation, grammar-switch count, subjective “does this still feel like speaking.”

Measure grammar, not whether a slot name is clumsy. Tasks must deliberately cross slots, or fill-in will always look sufficient.

Where it stops holding

Tasks that were forms to begin with (book a flight, expense a receipt, change a password) are not sacrificing openness; open input is the detour. A purely exploratory product with no preset task family, fully templatized, recaptures too far. A hybrid that declares “templates for the common, the box for the rest,” with artifacts importable either way, confines the sacrifice to the common path. Slot-filling on voice is especially awkward: speech has no cells, and forced recapture drives people back to a screen.

Applying it

• Decide whether this thing is really being recaptured. If yes, design it as a form and stop using “ask anything” as the lead copy. If not, templates can only be optional thin scaffolds, and what gets sent is still the natural sentence.
• Offer reversible conversion: values in cells can be assembled into a readable prompt, and that sentence can be split back into cells. A change of mind mid-way should not mean starting over.
• Leave a free-sentence position for cross-slot relations that will not be ignored, weighted in generation with the slots.
• Check: a request with “first… then only for…” on the main path. If the UI forces that sentence into unrelated cells and cannot assemble it back, openness has been recaptured. Then listen to what people call the entry — “ask it” or “fill the form.” When the name changes, the grammar has changed.

Related

• Same group: L2.03.1 Templates lower the cost of getting started · L2.03.2 Templates constrain the range of what users will express · L2.03.3 Templates must be editable, not only fillable · L2.03.5 Placeholder content in templates is kept verbatim; empty slots must be unskippable · L2.03.6 Templates encode the designer’s task assumptions and actively mislead when those do not match real work · L2.03.7 A template that can only be applied whole, with its structure invisible, teaches nothing
• Nearby: L2.01 Openness of Natural-Language Commands and Its Cost · L2.04 Complementarity of Parameter Controls and Natural Language · L2.06 Granularity of Control
• Search terms: open-to-form conversion · delayed binding · form versus NL