When motion, rendering, audio, or bandwidth fails, a conversational avatar should descend through explicit modes while preserving task state and accessible controls.

Read guide

02 / Inside the interface

Open the
black box.

Real source captures. Different layers of the same problem. Select a view and look at what the interface makes visible.

FIELD CAPTURE / Ubisoft2024-03-19

Official prototype screenshot · Original source ↗ · source credited

What to notice

Read the scene, not just the sentence.

The mission map, character, and dialogue share a screen. Notice how much context exists outside the chat transcript.

Narrative constraints → dialogue → world state

Read the full case ↗Source material, not our product test.
Public reuse requires owner review.

03 / Explore the connected field

No interface
is an island.

Good conversation connects design, knowledge, operations, and human judgment. Rotate the sphere. Pick a discipline. Follow the connection.

Connected disciplines.
One accountable system.

Enter the interactive lab ↗This is a conceptual map, not geographic data.

Conversation universe · conceptual map

Drag to rotate · Arrow keys to explore

01 / Foundations

Start with the job, the boundary, and the state the person can inspect.

Explore Foundations guides

NEW / FIELDWORK EDITION

The details
that make it work.

20 deeper guides for the difficult middle: correcting a spoken identifier, sharing an avatar, testing a non-answer, and tracing the wait between turns.

Choose your next five reads ↗

INTERACTIVE / VOICE LATENCYWhere does
the pause go?

Change six stage budgets. Inspect the first-audio critical path.

Illustrative model—not a live benchmark ↗

04 / Your next three reads

Start with
the job.

You don’t need to read everything. Pick the work in front of you and follow a connected route through the guide.

Turn a promising demo into a dependable system.

Choose the architecture

Decide what should be a flow, a retrieval step, or a bounded agent.

Make state explicit

Give every transition a precondition and a failure path.

Control what changes

Keep authorization outside the model’s persuasive language.

Choose a discipline

A field guide.
Not a feed.

Six connected desks. Curated starting points, concrete review artifacts, and deeper routes for the work behind a useful conversation.

01 / Foundations

Define the system.

A useful conversation starts with a job, a boundary, and an honest promise.

7 guides 02 / Conversation design

Make the next turn clear.

Design correction, uncertainty, waiting and evidence into the interaction.

9 guides 03 / Knowledge & actions

Ground it. Bound it.

Connect knowledge and tools without turning retrieved text into authority.

20 guides 04 / Evaluation & operations

Make failure observable.

Replace demo confidence with reviewable evidence, release controls and recovery.

16 guides 05 / Handoff & service

Keep a human path open.

Automation is part of a service with queues, people, capacity and consequences.

7 guides 06 / Embodied systems

Give presence a purpose.

Build the voice, avatar and world around the task—not the other way around.

17 guides

implementation
and operations guides

RESOURCE EDITION / SEPTEMBER 2026

The service behind the demo

Prepare for the
turn that goes wrong.

What happens when a retry arrives twice, evidence belongs to another customer, a voice connection drops, or an action needs a human to recover it? The resource collection follows those boundaries into the details.

Explore the resource collection ↗Inspect its coverage and limits ↗

Read → inspect → apply