01 · Problem
Learners often don’t know their question yet.
After a surprising result, learners want to ask about it before they can phrase it. Saying “wait, why?” is easier than typing a question.
The hard part was making it feel like a real conversation. A guide that lags or talks over you would undercut a tool about questioning AI.
02 · Conversation principles
The guide follows how people really talk.
People follow unwritten rules when they talk. I used these rules as the brief and wrote each one into the guide’s instructions.
Grice’s maximsConversation analysisTurn-takingRepairPoliteness
GriceQuantity
Say only what’s needed.
In the guideAnswers are 2 to 4 short sentences, easy to hold in your head.
GriceQuality
Only say what you can back up.
In the guideReal-world claims are hedged. Numbers from the demo are never called real statistics.
GriceRelation
Stay on topic.
In the guideOff-topic questions get a friendly nudge back to AI bias.
GriceManner
Be clear and brief.
In the guideEveryday words and real examples. No lists or code read aloud.
Sacks et al.Turn-taking
Wait for your turn.
In the guideIt waits for a pause, hands the turn back with a question, and stops when the learner speaks.
Schegloff and SacksAdjacency pairs
A question expects an answer.
In the guideThe greeting ends with a question, so learners know it’s their turn.
Schegloff et al.Repair
Fix mishearing quickly.
In the guideIt asks once: “Could you say that again?” Coughs and noise get no reply.
Clark and BrennanGrounding
Check you understood each other.
In the guideA label shows when it’s listening or thinking, and every line appears as text.
Brown and LevinsonFace
Protect people’s self-image.
In the guideThe guide takes the blame when something fails and never tells anyone off.
The greeting
First visit · welcome and introduction“Hi, I’m Northstar. Curious whether AI can be biased? You can ask me anything, or run an experiment to watch it happen. What sounds good?”
Returning · welcome back“Welcome back. Ask me anything, or run an experiment. Where do you want to start?”
New learners hear who is talking and what they can do. Returning learners just get a welcome back. Both end with a question and use the learner’s name if they’ve set one.
03 · Turn-taking
Designing the pause.
Most of what makes a voice feel human happens between turns. I designed three moments: when the guide starts speaking, how it hands the turn back, and how quickly it stops when the learner cuts in.
Listening
Waits for a natural pause.
Backup voice: 0.7 s of silence ends a turn. Sounds under 0.25 s are ignored.
On screen“Speak naturally. I’ll respond when you pause.”
Speaking
Stops talking the moment the learner speaks.
On screen“You can interrupt me anytime.”
Muted
Stops listening until the mic is back on.
On screen“Your microphone is off.”
04 · Error recovery
Planning for when it breaks.
Voice breaks more often than text. I planned a response for each failure, from a cough to a full outage, so the learner always knows what happens next.
MicrophoneA cough or background noise
Reacting to a cough would interrupt more than it helps.
Learner getsNothing. It keeps listening.
MicrophoneSpeech it couldn’t understand
The guide takes the blame for mishearing.
Learner getsOne short request to say it again.
ConversationAn off-topic question
Keeps the lesson going without telling anyone off.
Learner gets“That’s outside my focus,” and a way back to the lesson.
ServiceThe service is busy
Says what’s happening and what to do next.
Learner gets“Give me a second to catch up, then ask me again.”
ServiceReal-time voice can’t connect
The conversation keeps going.
Learner getsThe older, slower voice takes over on its own.
ServiceNo voice service at all
Experiments still work, since they don’t need AI.
Learner gets“The guide is resting, try again in a bit.”
An example
LearnerSo can a hiring AI be unfair even if nobody meant it to be?
NorthstarIt can. If it learns from past hiring decisions, it can pick up the same patterns, including unfair ones. Want to try the hiring experiment and see who gets left out?
LearnerWait, what do you mean by patterns?
Interrupts
NorthstarThings like which schools or neighbourhoods past hires came from. The AI can treat those as signs of a good candidate. What do you think that means for everyone else?
LearnerWhat’s the weather tomorrow?
Off-topic
NorthstarThat’s outside my focus, I only know about AI bias. Want to see how a loan AI decides who gets approved?
05 · Iterations
Five versions to get it right.
Record, transcribe, answer, speak
Couldn’t be interrupted, and every turn used build credits.
The browser’s free speech tools
A robotic voice that cut people off after 0.35 s of silence.
Gemini’s free tier
Often busy, and its backup voice made the guide interrupt itself.
Claude writes the answers
Better answers, but still slow to arrive.
A real-time voice agent
One live connection. Fast, natural, and it stops when you speak.
06 · Development
From four steps to one live connection.
I compared four real-time voice services. The one I chose had to be free, work without logins, and keep Claude writing the answers.
| Option | Cost for a demo | Decision |
|---|
| Deepgram Voice Agent | $200 free starter credit, no card, and Claude available inside it | Chosen |
| OpenAI Realtime | Paid only | No |
| ElevenLabs Agents | 15 free minutes a month, then paid, with the language model billed separately | No |
| Gemini Live | Free, but on the same busy free tier as versions 3 and 4 | Shelved |
Before and after
07 · Voice states
An orb that moves with the voice.
The orb shows whether the guide is listening, thinking or speaking. It grows and glows with the voice, rising on each word and settling in the pauses.
ListeningA calm drift that grows as you talk
ThinkingA faster swirl while the answer is written
SpeakingSwells with each phrase
These orbs follow a simulated voice. With reduced motion turned on, they stay still.
08 · Reflection
What I’d take forward.
Design the architecture with the conversation
I spent four versions adjusting timings, retries and wording when the real problem was a pipeline that could only do one thing at a time. Next time I’d decide how turns, interruptions and failures should feel before choosing any technology.
Test with real conversations
So far I’ve tested the guide in my own sessions. Next I’d run sessions with learners, with their consent, and review where turns broke down or answers missed the point.
Design every error in detail
With more time, I’d design each failure as carefully as the main flow: write the exact message for every error case, decide what the screen shows, and make sure the learner always knows what to do next.
Next challenges
| Challenge | Why it matters |
|---|
| A classroom full of voices | Other people talking nearby could interrupt the guide or end a turn early. |
| Other languages | The guide only speaks English. Each new language needs its own rules, greeting and repair lines. |
| Microphone consent | It should be obvious when the mic is on, what is sent, and that nothing is kept. |
| Keyboard and screen readers | Voice mode should work fully without a mouse or sight, beyond the transcript. |