Chapter 8 - Layer 3: Dialog Interface (What the AI Implements)
Last updated 2026-07-13. Explore the AI-side dialog interface through dialog-layer prompts and worked transcripts, with CSAR as the human-side protocol for clarification, synthesis, action, and reflection.
Companion activity
Use the dialog worked example or dialog prompts to repair one request where the literal ask and the underlying need diverge.
What this chapter argues
The dialog interface is a genuine brain region, not just an input/output channel. The AI implements five concrete disciplines here: intent modeling (reading what was meant, not just what was said), faithful response generation (showing the work, and saying "I don't know" instead of offering a confident guess), clarification offers (asking one sharp question rather than five generic ones), surfacing assumptions, and dialog repair when grounding fails. These are AI-side disciplines, not pointers to other books.
Key sources
Chapter 8 cites in the print book: D11, D35, D37. Entries below are the sources this page draws on; some are related literature cited elsewhere in the book rather than in this chapter.
- D3 - Correa, F. (2024). Dialog Engineering: AI as Your Research Assistant. medium.com/@fabioc - the dialog-discipline essay that anchors the human-side counterpart.
Primary-source links
- Correa, "Dialog Engineering: AI as Your Research Assistant" (October 2024) - the article that named the discipline and introduced the human-side CSAR protocol.
Chapter contents
- The dialog layer reframed as a brain region (not just an I/O channel).
- Five AI-side disciplines (intent modeling, faithful response generation, clarification offers, surfacing assumptions, dialog repair).
- How the AI-side disciplines complement the human-side CSAR protocol.
- The interface between the dialog layer and the working-memory layer: how clarification interrupts the loop appropriately.
Related resources
The dialog-layer prompts page hosts copy-paste-ready templates for intent-clarification, faithful-response framing, assumption-surfacing, and dialog-repair openers. A dialog worked-example shows a full intent-clarification dialog start to finish.