synthesizing-discovery-researchlisted

# Synthesizing Discovery Research ## Overview Synthesis turns a pile of raw qualitative data into a small set of **evidence-linked insights** the team can act on. The discipline that matters most when AI assists: **every insight must trace back to something a real person actually said or did.** ## AI guardrails (read first) When using AI to code or summarize research: - **Never invent quotes.** ~8% of LLM-generated "quotes" can't be found in the source transcript — often subtle rewordings that distort meaning. Verify every quote against the raw text before it reaches a deck. - **No synthetic users / synthetic findings.** AI participants are sycophantic, one-dimensional, and report idealized behavior. Do not present AI output as real-user findings. - **Treat AI clusters/themes as a first draft to verify**, not as conclusions. - **Seek disconfirming evidence** — ask the model "what in the data contradicts this theme?" - **Cite the source** for every nugget; strip PII. ## Methods **Affinity mapping (KJ method)** — for making sense of a large pile fast: 1. One observation/quote per note. 2. Each person writes notes independently first (5–10 min) to avoid groupthink; cluster in silence. 3. Cluster by **meaning, not shared keywords**; let labels emerge. 4. Keep a **focus question** visible (e.g., "Why aren't users logging in?"). 5. Don't discard outliers — small clusters hold value. **Thematic analysis (Braun & Clarke, 6 phases)** — for rigor: familiarize → code (descriptiv