M1.06.3strategy shift after consecutive failuresdesignresearch

Consecutive failures must change strategy, not repeat the same prompt

Aliases: change the move after two misses · don't re-ask the same prompt · repair strategy escalation

What it is

A place name misses two turns running. If turn three is still the same “please say that again,” the user will send the same failed form — open street name, same accent, same noise. Strategy shift after consecutive failures requires that the second or third miss change the move, not say the same move again. The change is structural: collapse to yes/no, switch to spelling, hand off to a screen or keypad — not a synonym for “I didn’t catch that.” Synonym wording treats boredom. Strategy shift treats a path that has already been shown not to work.

Why it happens

The same prompt shapes the same class of answer. Open “where to” induces a full street name; asking that sentence again after the street name failed leaves the prior and the recognition conditions unchanged, so the expected output is still that high-open slot. Escalation cuts the hypothesis space: city versus district first, then a closed choice on what remains, then an initial or a list to tap. Each step admits that this channel plus this slot type has already given evidence; another sample adds no information. Consecutive failure also spends patience in common ground: people start to doubt the skill exists at all, not this one acoustic take. So a third open chase abandons more than “still wrong, but the method already changed.” Changing channel is also strategy, not surrender — it proves this beat of voice is a poor fit for this slot, not that the whole task should die.

Studying it

Build two paths on the same hard slot: the same prompt three times versus a structural change from the second try (open → closed alternative → on-screen list). Dependents: success on the third attempt, abandonment before the third, whether people start slowing down or shouting letter by letter. Independents: slot openness, noise, whether a second channel exists.

On logs, plot “consecutive recognition failures on the same slot” per skill, and the fraction of the nth try that still uses the original prompt. If success stops rising for n≥2, repetition has saturated. Wizard-of-Oz can test the escalation ladder itself with a deliberately degraded recogniser, so acoustic improvement and strategy improvement are not tangled. Do not count a synonymous re-prompt as a strategy change — if the syntactic expectation is unchanged, it is the same move.

Where it stops holding

A first miss is often a slip or a brief noise; escalating immediately to spelling over-punishes and makes the system look fragile. Consecutive failure on a closed skill (on/off) is more likely a microphone or device problem; strategy should lean toward confirming the channel or closing, not inventing a weirder way to ask on/off. On a device with no second channel the ladder still has to exist (shrink the set, ask an attribute); it cannot recite the same sentence until the user walks because there is nowhere to hand off. High-consequence slots must not quietly drop confirmation strength while escalating (two failed street names, then execute the nearest popular place). Children and older adults fail in different patterns; the ladder should be read per population, not as one shared “spell after two misses.”

Applying it

  • Write a failure ladder per open slot: try 1 may repeat the original structure; try 2 must change the expected output (closed choice, attribute filter, spelling); try 3 defaults to a second channel or a recoverable draft.
  • Each step on the ladder must make the new demand hearable (“Nanshan or Futian,” not “please say the destination again”), or people will still send an open street name.
  • Do not let a synonymous re-ask stand in for step 2. If that step still wants the same open value, strategy did not change.
  • How to check: force an open place name to fail twice. If the third system move still wants a full address, the ladder is not working. What should appear is a cut-down set, or a channel already handed off.

Related

  • Same group: M1.06.1 Users must be able to correct the part that was misunderstood · M1.06.2 Partial correction beats repeating the whole utterance
  • Nearby: M2.02 Open and Closed Questions · M2.04 Error-Recovery Wording · C7.17 Switching Between Voice and Keyboard
  • Search terms: strategy shift after consecutive failures · repair escalation · prompt adaptation

Cards in the same group

Quick Actions

Share

Share this page

ios_share

https://hci.top/en/handbook/M1.06.3