UX Researcher

1. 1 Tag by theme Claude Haiku 4.5 720ms

   Open-coded 12 transcripts into 34 codes. Four clusters carry most of the weight; the rest are singletons.

2. 2 Synthesise findings Claude Opus 4.8 1520ms

   Nine of twelve stalled at the workspace-invite step. They all said a version of "I don’t know who to add yet" — that is one finding, not nine.

3. 3 Audit the evidence GPT-5 980ms

   Checked every claim against a timestamped quote. One finding — "users want templates" — rests on a single participant. Demoting it to a signal.

Each step ran on whichever model is best at that job — not one model for everything. The worker is the workflow.

• Read: session transcripts, study protocols and prior findings — with participant PII already redacted at ingest
• Write: findings documents and the evidence log only; it never edits or deletes a source transcript
• Escalate: the researcher when a claim rests on a single participant, or when two sessions directly contradict each other
• Never: contact a participant, or surface anything that could re-identify one
• Never: paraphrase a quote it uses as evidence — every claim links to a timestamped verbatim

• 4 findings, each traced to at least 3 timestamped quotes.
• 1 demoted to a signal — single participant, flagged not dropped.
• Zero paraphrased quotes. Every line of evidence is verbatim.

private rules were taught in one workspace and never leave it. general rules are true about the work, so every copy of this worker gets them.