V5 → V6
Clearer recovery paths
Typing states, drawer support and in-context grammar improved lesson progression before Suggestion addressed pronunciation failure.
02 of 06 · Core product
Designing conversation as a language lesson.
I developed the core AI conversation model, then iterated using observed frustration, funnel behaviour and product versions until Suggestion helped learners recover from pronunciation failure.
At a glance
- Problem
- The conversation interface could identify incorrect responses, but repeated pronunciation failures created frustration and lesson abandonment.
- Key insight
- Learners did not only need correction. They needed contextual support that explained why an answer failed and how to continue.
- Decision
- Design the conversation as a structured learning system and introduce Suggestion as an in-context recovery mechanism.
- Result
- Suggestion increased lesson completion by 25% by helping learners recover without removing the learning challenge.
01 Challenge
Turning AI conversation into a teachable lesson.
Conversation was the core of the Kaizen experience, but a conventional messaging interface would not provide enough support for someone learning a new language.
Before defining the interaction model, I reviewed messaging and language-learning products to understand established behaviours, including message hierarchy, audio controls, conversation pacing and how supporting information was presented. Competitors provided useful interface patterns, but none were solving the same problem: creating a conversation that felt natural while also functioning as a structured AI-powered lesson.
02 Interaction model
Four learner needs, one conversation
Rather than treating each control as a separate feature, I organised the lesson around what learners needed at each moment — from understanding the AI's state to recovering when confidence broke down.
Understand what the AI is doing
AI responding
Clear composing and speaking states so the lesson feels active.
Hear and interpret the response
Audio playback
Animated feedback when the tutor speaks.
Reduced-speed pronunciation
Normal and slowed playback when a phrase is unclear.
Translation
English support revealed on demand, not shown by default.
Character switching
Romaji and hiragana so reading difficulty could adapt.
Access additional support when needed
In-context grammar
Drawer cards and prompts without leaving the conversation.
Recover from errors without losing the lesson
Errors and retries
Clear error states with retry actions.
Suggestion
Contextual pronunciation support after repeated failure — the key recovery decision.
03 Evidence
What testing and behaviour revealed.
Through guerrilla testing and recorded sessions, I found a recurring issue: users often believed they were saying the right word, but still received errors. After repeated failed attempts, frustration increased and lesson drop-off rose sharply. Correction alone was not enough support.
Users needed help recovering from mistakes, not just being told they were wrong.
Mizuki lessons · iOS · Last 7 days
72.0%
Average lesson completion
Key drop-off finding
Most learners sent messages successfully, but completion varied —mizuki_1 dropped to 59.16% overall, signalling friction after repeated failure rather than at lesson entry.
Lesson progression funnel
- Chat lesson started
- Message sent
- Lesson completed
04 Iterations
From static lessons to interactive learning loops.
| Version | What failed | What changed | What evidence showed |
|---|---|---|---|
| Version 5 | Limited interaction states and weak recovery when learners became stuck. | Baseline conversation model with differentiated messages and core controls. | Funnels showed drop-off after repeated pronunciation failure. |
| Version 6 | Learners needed clearer feedback and support without leaving the lesson. | Typing states, drawer support, grammar prompts and clearer options. | Testing showed improved responsiveness; pronunciation recovery remained unresolved. |
| Later versions | Supporting content sat too far from the instruction. | Controls moved closer to instructions; grammar accessible in context. | Mixpanel compared version progression from entry to completion. |
05 Suggestion
Helping users recover when pronunciation failed
To reduce frustration, I introduced Suggestion, a support card that broke pronunciation into simpler syllables and allowed users to hear the word more slowly. The goal was to preserve challenge while giving users a way through moments of failure. The design was triggered after repeated errors, directly inside the lesson flow.
- Reduced dead ends in the lesson
- Gave users a clearer path to success
- Supported confidence without removing learning effort
- Kept help contextual rather than hidden elsewhere
Testing revealed repeated pronunciation failure. Correction alone was not enough. Suggestion broke pronunciation into manageable support while keeping the learner inside the lesson — increasing lesson completion by 25%.
06 Delivery
Defining the behaviour for development
I designed the individual message states, audio behaviours, errors and character-set variations, then documented how each component should respond throughout the conversation.
I also created an interactive prototype to communicate timing, audio feedback, animation and transitions to the engineering team—details that could not be understood fully from static screens alone.
- Differentiated chat bubbles and message states
- Typing and audio feedback to make AI feel responsive
- Translation controls
- Character-set support such as romaji and hiragana
- Instructional cards and correction states
- Prototypes and documentation for developer handoff
Measured
Behaviour-led iteration
Mixpanel funnels compared lesson versions so design changes could be assessed against real completion behaviour.
Foundation
Connected product work
The conversation model informed later areas including practice and retention mechanics. See retention case study.
07 Results and learning
Failure is not the same as unrecoverable difficulty.
Suggestion worked because it distinguished between failure and recoverable difficulty. The lesson remained challenging, but learners had a clearer route forward when confidence began to break down.
The conversation model became the interaction foundation for later product areas, with measurable improvement tied specifically to in-lesson recovery.
Improving error recovery helped learners progress through conversation lessons, but research revealed a more foundational barrier: many beginners lacked confidence with Japanese writing.
Continue to Japanese writing system