How the guide talks, listens and recovers. For the short version, see the main Voice Agent page.
| Principle | What it says | How I applied it |
|---|---|---|
| Quantity · Grice's maxims | Say as much as the moment needs, and no more | Answers are 2 to 4 short sentences. A listener hears each sentence once and has to hold the whole answer in their head |
| Quality · Grice's maxims | Only say what you have evidence for | Real-world claims are hedged with "can", "often" or "researchers have found". Numbers from the teaching model are never presented as real statistics |
| Relation · Grice's maxims | Stay relevant to the conversation | The guide only talks about AI bias. Off-topic questions get a friendly redirect back to the lesson |
| Manner · Grice's maxims | Be clear, brief and orderly | Everyday words, examples from hiring, loans, healthcare and school admissions, and no lists, symbols or code read aloud |
| Turn-taking · Sacks, Schegloff and Jefferson | People signal when their turn is ending, and speakers rarely overlap for long | The guide waits for a natural pause, usually ends with a short question to hand the turn back, and stops as soon as the learner starts talking |
| Adjacency pairs · Schegloff and Sacks | Some turns expect a matching reply | The greeting ends with a question, so the learner knows exactly what is expected of them |
| Repair · Schegloff, Jefferson and Sacks | Conversations recover from mishearing quickly, usually with a short request to repeat | One short sentence asks the learner to say it again. Noise and coughs get no response at all |
| Grounding · Clark and Brennan | Speakers keep checking they've understood each other | The state label shows that the guide is listening or thinking, and every line appears as text |
| Face · Brown and Levinson | People protect each other's self-image | When something goes wrong, the guide takes the blame. Off-topic questions are redirected without telling anyone off |
The greeting is a fixed script, so every learner hears exactly the same first words.
| State | What the guide does | On screen |
|---|---|---|
| Listening | Waits for a natural pause | "Speak naturally. I'll respond when you pause." |
| Thinking | Writes the answer | "One moment…" |
| Speaking | Answers, and stops as soon as the learner starts talking | "You can interrupt me anytime." |
| Muted | Keeps the conversation open but stops listening | "Your microphone is off." |
| Waiting for permission | Shown if the microphone isn't allowed within about a second | "Allow microphone access to start talking." |
| What happened | What the learner gets | Why |
|---|---|---|
| Only noise, a cough or an echo | Nothing. The guide quietly keeps listening | Apologising for a cough would interrupt more than it helps |
| Speech it couldn't make out | One short sentence asking them to repeat it | The guide takes the blame, so the learner doesn't feel they said it wrong |
| The service is busy | "Give me a second to catch up, then ask me again." | Says what's happening and what to do next |
| An off-topic question | A friendly "that's outside my focus" and a suggestion to get back on track | Keeps the lesson going without telling anyone off |
| The real-time voice can't connect | The older, slower voice takes over on its own | The conversation keeps going |
| No voice service at all | "The guide is resting, try again in a bit." | The experiments still work, because they don't need AI |
The orb's size, swirl and glow follow how loud the voice is, smoothed like breath: it rises in about 0.08 seconds and falls over about 0.35 seconds. Thinking breathes slowly on its own. The state is always written underneath too, and with reduced motion turned on the orb stays still.