ai-interaction-patterns

Start with the product-type table: chat assistants need Open Input, Suggestions, Follow-ups, Memory, and Disclosure as essentials, with Gallery, Voice & Tone, and Branches as optional additions. Then consult the full reference for design guidance on each pattern.

The Trust Builders category covers Disclosure, Caveat, Consent, Data Ownership, Watermark, Footprints, and Incognito Mode. For high-stakes domains like healthcare or finance, combine Citations, Verification, Disclosure, and Caveat together rather than relying on caveats alone.

Memory should never be a black box: show users when memories are added, and allow viewing, editing, and deletion. Choose between global, scoped, or ephemeral memory scopes, and offer an incognito mode for non-persistent sessions.

Show an Action Plan before any complex, expensive, or multi-step task so users can confirm or modify steps before resources are consumed. Use contractual mode requiring explicit approval for irreversible or high-impact operations.

Key anti-patterns include photorealistic avatars that overpromise capability, hidden memory, missing stop controls during generation, relying solely on caveats for safety, and overwriting user work without verification. Each erodes user trust in distinct ways.