User Stories: Curator pilot MVP (trustworthy session loop)
Detailed P0 stories below; P1s summarized at the end. Derived from 13-deliver-prd.md.
US-1 — Speaker-separated transcript of a consented session
| Field | Value |
|---|---|
| ID | US-1 |
| Persona | Practicing coach |
| Priority | P0 |
| Epic | Session intake & transcription |
| Estimate | M |
As a coach, I want a speaker-separated transcript of a consented session, so that I have an accurate, attributable base for analysis.
Acceptance Criteria
AC-1: Given a session with recorded consent, When I upload/connect the recording, Then a transcript is produced with client vs coach turns labeled. AC-2: Given no consent is recorded, When I try to process, Then the system refuses transcription and tells me why. AC-3: Given overlapping speech/low audio quality, When diarization is uncertain, Then affected turns are flagged low-confidence rather than mislabeled silently.
Technical Notes: integrate ASR/diarization vendor; target accuracy in metric 10. Out of Scope: real-time/live transcription.
US-2 — Reproducible, evidence-quoted analysis
| Field | Value |
|---|---|
| ID | US-2 |
| Persona | Coach |
| Priority | P0 |
| Epic | Analysis pipeline |
| Estimate | XL |
As a coach, I want a structured analysis where every conclusion has a transcript quote, so that I can trust it enough to act on and send.
Acceptance Criteria
AC-1: Given a depersonalised transcript, When analysis runs, Then I get topic, insights, commitments, and next steps, each with ≥1 transcript quote. AC-2: Given the same session, When analysis is re-run, Then extractions agree within the defined noise band (reproducibility). AC-3: Given a coach gold-labeled sample, When compared, Then F1 ≥ 0.70 on attribution/extraction.
Technical Notes: multi-step pipeline (ADR-001); not a single prompt. Out of Scope: clinical interpretation.
US-3 — Editable, coach-sent client summary
| Field | Value |
|---|---|
| ID | US-3 |
| Persona | Coach |
| Priority | P0 |
| Epic | Follow-up delivery |
| Estimate | M |
As a coach, I want an editable draft client summary that I explicitly send, so that I follow up fast while staying the author.
Acceptance Criteria
AC-1: Given a completed analysis, When I open the draft, Then I can edit it freely before sending. AC-2: Given an unedited or edited draft, When I have not pressed send, Then nothing is delivered to the client (no auto-send). AC-3: Given I press send, When delivery succeeds, Then the summary is marked sent and logged.
Out of Scope: automated send scheduling.
US-4 — Depersonalisation before AI (P0)
As a coach, I want identifiers removed before any external LLM call, so that I protect my client's data. (ACs in 15-deliver-acceptance-criteria.md, AC-5/error states; PII-leak audit = 0.)
P1 stories (summary — detailed later)
- US-5 Client dossier: accumulates insights/commitments + status across sessions; opens pre-session.
- US-6 Out-of-scope alert: flags clinical/crisis signals to the coach; system does not act; coach decides referral.
Open Questions
- [ ] Reproducibility threshold value (AC-3 of US-2) — confirm with coaches.
- [ ] Which delivery channels for "send" (email + messenger)?