SwissDnaCode — refinement summary

Most of Workshop 2's scope holds. The changes below are the items that moved between MVP, Vision, and Out — or that were explicitly added to the MVP this round.

Everything else from Workshop 2 stands: two-bot MVP, web-first with native-app architecture path, three roles, active-substance-based medication, no supplement sales on platform, no data migration of the existing 10 clients.

A serious incident this week sharpened the urgency to migrate. The current ChatGPT-based "Identity Code" bot started referring to the platform's own creator when answering a customer — describing her as the author of the DNA analyses and books — which confused the customer and damaged trust ("how much of the rest is even correct?").

Workshop 2's stance ("data stays in Switzerland") softens slightly. The position is now nuanced: patient data stays in Switzerland as the central USP; other / form data may sit elsewhere in Europe if it delivers real value and the cost relation is favourable. A rigid "Swiss-only" rule across every byte risks tripping the project up without real customer benefit — but the sales story still needs "data in Switzerland" to land cleanly.

Cost is still a black box — concrete figures emerge from the LLM and cloud research, and that calculation now becomes urgent because it drives the subscription pricing. Workshop 2 dropped credits in favour of fixed-price products; Workshop 3 sketches the first concrete tier structure on top of that, plus how usage caps are drawn.

Three tiers after the initial 3-month coaching

Workshop 2 captured consent as a mandatory statement at onboarding. Workshop 3 refines this: legally, a "No" option must be available. A pre-ticked, non-removable share-with-doctor checkbox is not viable. The flow now sketches what happens when a customer declines — and how they can change their mind later.

Workshop 2 set the principle that both bots ship in the MVP — Workshop 3 changes the build sequence. Rather than building the doctor bot first and the customer bot second, both are built in parallel and tested simultaneously: Nadine tests the doctor bot, Dante registers as a customer and tests the customer bot. This surfaces where restrictions are needed faster, and lets the team later swap roles to walk a full "hit a restriction → book an appointment with the doctor" loop.

Reconfirmed: same data basis, two interfaces — customer bot reduced and safer, doctor bot fully medical. Customer chat history stays private; doctor history contains only Dr. Farkas's questions. Patient uploads attach automatically once consent exists.

Workshop 2's principle holds: recommendations go by active ingredients, not product names — DACH markets use different preparations and SNP analyses map cleanly to substances. Workshop 3 adds a new option on top: a medication check via the Progenom lab against the ~500–800 medications actually tested in the DNA analysis, sidestepping AI hallucination on this specific question.

If Progenom is chosen as the lab, Mendelio can wire in the internal shop from day one — direct product orders, status / workflow, automatic import of findings into customer accounts. The infrastructure already exists; reusing it eases organisational load. If another lab is chosen, the team still provides internal organisational support, just without the API integration.

Business framing (thought, not decision): the DNA test itself need not be a high-margin product. The value sits in the app — selling the test at a small margin to enable the platform is a valid commercial choice.

Workshop 2's plan stands: a negative-prompt layer (an explicit list of restricted answers) is layered over the LLM so behaviour stays under control even when the model updates. Workshop 3 adds a refusal model for off-topic questions and reconfirms storage / categorisation.

Storage & categorisation reconfirmed: store questions, answers and summaries, tagged by category (sleep, weight, stress, sport — medical vs. lifestyle); the AI reads prior summaries first for long-term answer stability. Categories to be defined jointly from a clinical perspective.

Workshop 3 sharpens the data-protection rule that sits under the role architecture: the super-admin / admin must not see patient data; the doctor may see it but may not make system-relevant settings. Both founders can hold super-admin of the platform and doctor-admin of the practice — these are separate hats. Future employee or reseller roles always carry fewer permissions than admin or doctor-admin.

Two hats, same person: both founders can be platform super-admin and practice doctor-admin simultaneously. The architecture distinguishes hats, not people.

Doctor menu reconfirmed: patient list with search, patient detail as the central file — the three earlier tabs (patient details, doctor bot, all findings) collapse into the unified card view. Each role has its own settings. Customer menu: Dashboard, First Steps (incl. doctor data-processing consent), DNA chatbot, My Findings, Settings, plus Notifications.

A direct doctor-to-patient chat was raised and deprioritised. The reasoning: it creates tight time-binding obligations and a new tool to manage — neither fits V1. If pursued, it lands in the Vision / VIP / premium tier and is constrained via bookable time slots so the doctor is only reachable in defined windows, like an online calendar.

Why deprioritised: an always-on chat with the doctor creates time-binding obligations and a new tool to manage. The Vision-only treatment keeps V1 lean.

Workshop 2's four KPIs stand without change. Reconfirmed here so the next phase doesn't re-litigate them.

Staged feature release reconfirmed: features unlock per user on a time basis from registration — price stays the same; features arrive over time. A positive "getting more over time" feeling. Nadine's opening comments (incl. cancer patients) to be added to this section in the next prototype version.

Workshop 2's web-first stance holds. Workshop 3 sharpens the wording on the native-app path: the fastest store route is a downloadable "window frame" (a wrapper) so content can be updated server-side without re-submitting to the stores. The customer downloads an app and just logs in — no URL to type.

Main point: the customer downloads an app and just logs in — they never type a URL. Three environments (production, customer testing, programming sandbox) live under the same architecture.

This summary captures the third workshop's decisions, refinements, and open threads. Prototype extension continues; LLM and cloud-cost research is the central gating task because everything downstream depends on it.

Open points — still to decide

Prototype work continues from Monday; LLM and cloud research starts next week and is the central, gating task — costs define pricing downstream. The Monday workshop is not needed; the weekly Friday meeting is kept for next week, expecting prototype results and possibly LLM decisions or installations.