Stop Writing Tone Instructions. Layer Them.
summary
A production-tested architecture breaks voice into four executable layers: immutable identity (rules like disclose you are an AI in first response and never claim physical presence), situational mode that loads role and soft-context notes, example-anchored voice for phrasing, and a deterministic post-generation veto that blocks lies—e.g. replacing a brittle stack of 24 different system prompts with one assembler and a veto that prevents offering nonexistent dates or falsely “matching” missing people.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:02Single system prompts fail
- 2:29Four-layer architecture
- 5:15Layer one: immutable identity
- 8:23Layer two: situational mode
- 11:53Layer three: anchored voice
- 13:36Layer four post-generation veto
- 16:24Veto is most critical
- 17:50Multi-tenant systems
- 18:30Instructions vs permission
- 19:15Lessons learned
related talks
17mBuilding Effective Voice Agents
13mStepWrite: Adaptive Planning for Speech-Driven Text Generation
28mWhat’s Needed in Tech to bring Voice to the Future
17mThe Voice-First AI Overlay: Designing Conversational Co-Pilots
57mDesigning Human-AI Experience
48mStop Fighting Your AI Engineering Prompts That Actually Work
55mDesigning for speech
17mReimagining Customer Experiences Integrating Voice into Digital Strategy