01
02
03
04
05
06
Connect Claude to your Gong or Fireflies transcript API so that every new sales call or user interview is automatically processed through a pipeline of specialist agents: one segments the conversation, one extracts Jobs-to-be-Done signals in structured format, one mines verbatim quotes, and one spots competitor mentions with switching context. All signals are stored in a Supabase table. A synthesis agent runs on schedule, aggregates signals from the last 90 days, and generates a product brief grounded entirely in actual customer language — no PM intuition required.
6–10 hours per quarterly product synthesis cycle; ongoing elimination of insight loss between calls and planning
Product
Claude
Gong Enterprise account with OAuth2 app credentials (client_id, client_secret) and api:calls:read scopes — OR Fireflies API key (simpler setup, no enterprise plan required); Supabase project URL and service role key; Claude API key; Python 3.10+ environment with LangGraph and the relevant transcript SDK; list of customer segments to track (SMB / Mid-Market / Enterprise)
Run the first synthesis on a batch of calls where you already know the key themes from memory. Verify the JTBD themes the agent surfaces match your intuition — if they do not, check whether the Segment Classifier is correctly isolating discovery content from demo content. Confirm verbatim quotes are preserved exactly with no paraphrasing in the signal store. After 3 months of operation, test an ad-hoc query: "How many enterprise customers mentioned [competitor] in the last 6 months?" and verify the answer is correct.
Gong and Fireflies API access requires their respective paid plans — check your current subscription tier before building. Transcript quality degrades with poor audio, heavy background noise, heavy accents, or domain-specific jargon the transcription model was not trained on. JTBD extraction accuracy improves significantly when your reps ask structured discovery questions — calls that are mostly demo with no discovery produce minimal signal. The synthesis is only as good as the signal volume: fewer than 20 calls in the 90-day window produces unreliable theme clusters. Privacy and compliance: customer call transcripts contain personally identifiable information, company details, and confidential business context — ensure your Supabase instance is compliant with GDPR, CCPA, or the data protection regulations applicable to your customers before storing verbatim quotes; obtain any consent required under your customer agreements.
Stack: Gong REST API or Fireflies GraphQL API (transcript retrieval), LangGraph (agent orchestration), Anthropic SDK (Claude Sonnet for all extraction and synthesis agents), Supabase (persistent signal store and ad-hoc querying), Python. The verbatim quote preservation step is critical — the value of the product brief comes from real customer language, not PM reinterpretation. Store the raw quote in the database; never let an agent paraphrase it before storage.