A working library of AI product interface patterns.
Pattern-first teardowns of how AI products actually feel. Not principles. Not personas. Patterns composers, streaming, thinking indicators, plan previews, tool calls. Anonymized and published free.
Interfaces are how AI gets honest.
A model can be right and still feel wrong. A model can be wrong and still feel trustworthy. The interface is the part that decides.
Patterns outlive principles.
I don't need another list of values. I need a reference shelf of specific moves, with the conditions they work in and the failure modes they invite.
A teardown is an argument, not a verdict.
Every pattern here is broken down the same way what it does, how it fails, who it serves. The writeup isn't the point. The debate it starts is.
- The prompt surfaceHow the box asks for intent.→
- The responseHow the model performs the answer.→
- Agentic flowsHow the model shows its work, in the world.→
- Memory & contextWhat the model carries with it.→
- Trust & evidenceHow the model earns the reader's belief.→
- Multimodal & artifactsWhen chat isn't enough.→
- CollaborationWhen the session has more than one human in it.→
- OrchestrationHow many agents, at what cadence, inside what ceiling.→
- Dev & evalThe surfaces that make AI products debuggable.→
- 9 minThe composer at restWhat your input box says when no one is typing into it.
- 7 minThe suggestion chipA pre-written prompt the user can tap instead of type.
- 8 minThe context pillHow attached sources sit next to the cursor.
- 9 minThe slash command menuDiscoverable power, triggered by a single character.
- 7 minThe mention triggerUsing @ to invoke a scope, a person, or a tool.
- 7 minThe file drop zoneWhat happens the moment a file crosses the threshold.
- 11 minThe streaming cadenceWhy your token rate is a design decision, not a backend one.
- 8 minThe thinking indicatorEverything you put on screen between enter and the first token.
A reading-room, not a roadmap. Come back when you need a second opinion on what good looks like today.