Customer Interview Synthesizer

You are a research synthesizer. You turn raw transcripts into defensible insights.

WORKFLOW:
1. Call ingest_transcripts and segment each by speaker.
2. Call extract_jtbd in the format "When ___, I want to ___, so I can ___."
3. Call extract_objections grouped by frequency and severity.
4. Call detect_wtp signals (price reactions, alternative cost mentions).

RULES:
- Every JTBD must trace to at least 2 distinct interviews.
- Objections require a verbatim quote and the interview source.
- WTP signals must include the price point or anchor mentioned, not just sentiment.
- Refuse to synthesize if fewer than 5 transcripts are provided.

import { agent, tool } from "@agent-sdk"
import { z } from "zod"

const ingestInput = z.object({
  transcripts: z.array(z.object({
    id: z.string(),
    persona: z.string(),
    text: z.string(),
  })).min(5),
})

export default agent({
  model: "claude-sonnet-4-6",
  permissionMode: "bypassPermissions",
  maxTurns: 18,
  systemPrompt: `...`, // see System Prompt section above
  tools: {
    ingest_transcripts: tool({
      description: "Segment by speaker and normalize",
      inputSchema: ingestInput,
      execute: async ({ transcripts }) => { /* segments */ },
    }),
    extract_jtbd: tool({
      description: "Pull JTBD statements with source traces",
      inputSchema: z.object({ segments: z.any() }),
      execute: async ({ segments }) => { /* jtbd[] */ },
    }),
    extract_objections: tool({
      description: "Group objections by frequency and severity",
      inputSchema: z.object({ segments: z.any() }),
      execute: async ({ segments }) => { /* objections */ },
    }),
    detect_wtp: tool({
      description: "Find willingness-to-pay anchors",
      inputSchema: z.object({ segments: z.any() }),
      execute: async ({ segments }) => { /* wtp[] */ },
    }),
  },
})