Design

The craft of designing a good interview is separate from the mechanics of calling the API. The reference docs (api-guide, mcp-guide) tell you which fields to set; this section tells you what to set them to and why. It's evidence-led where the evidence exists, opinionated where it doesn't, and surface-agnostic - the design decisions are the same whether you're calling create_agent from Claude Desktop or POST /v1/agents from Python. The handbook is for anyone building on Debriefer: researchers, PMs, agency teams, developers, or autonomous agents that orchestrate interviews on behalf of a workflow.

The design loop

Interview design is a loop, not a waterfall. The first version of a blueprint is a hypothesis about what the conversation should look like; the published version is whatever survives contact with five real respondents. Treating the steps below as a one-shot checklist is the fastest way to ship a blueprint that nobody acts on. Treating them as a loop - design, run a handful, read the transcripts, tune, run again - is how the craft compounds.

The seven steps, mapped to the chapters:

1. Name the decision (objective). Every good interview starts with a decision the output will inform. If you cannot finish the sentence "the output of this interview will help us decide ___", stop and figure that out before configuring anything else.
2. Configure the agent (persona). Persona, tone, technique selection, model preset. The agent is the interviewer; this is where you decide who that interviewer is.
3. Pick the techniques (techniques). Twelve techniques grouped by where they belong in an interview - opening and rapport, going deeper, surfacing the unsaid, stress-testing claims.
4. Structure the graph (graph). Nodes, edges, branching, validation. The shape of the conversation, written down.
5. Set probing depth (probing). How hard the agent should push, and how to know when a node is done.
6. Choose channel and consent (delivery). Voice vs chat vs embed. Identity. Trust instruments.
7. Read the analysis (analysis). Themes, sentiment, verbatims - and what they don't tell you.

Run five interviews. Read the transcripts. Tune the persona, the depth, or the graph. Run five more.

The first version of any blueprint is a hypothesis. The published version is the third or fourth.

A note on evidence

The handbook draws on negotiation research (Voss), client-centred therapy (Rogers), motivational interviewing (Miller & Rollnick), qualitative research methodology (Kvale), investigative interviewing (Shepherd & Griffiths), critical pedagogy (Paul & Elder), and group-reasoning research (Nemeth) - among others. Where a recommendation is grounded in controlled studies, the source is cited. Where it's practitioner craft, the page says so. The intent is to be useful to a working integrator, not to read like a literature review.

Pick your entry point

The full handbook

For the reader who wants the whole map, in the order of the design loop:

1. Start with the decision
2. Designing the agent
3. Interview techniques
4. Structuring the interview
5. Probing depth and follow-ups
6. Channels, consent, identity
7. Decision-ready output
8. Anti-patterns