Workshop 04 · Live prototype walkthrough
A shared record of the fourth workshop — the first live prototype walkthrough of V1 and Vision, a review of Nadine's onboarding-workflow document and Dante's bot-behaviour document, and the operational decisions that follow from real usage feedback with the current bot's customers.
What we settled. Testing effort focuses on V1 — Vision stays documented and evolves step by step. Nadine's onboarding-workflow document and Dante's bot-behaviour document (traffic-light restriction levels 1–3) are treated as a strong base and rolled into the prototype for the next meeting. Certification is not pursued now; Memberspot stays for the academy in V1 (a link, optionally an API integration later); the full academy is Vision.
UI direction locked. The doctor bot gets a quick-access field/icon directly under the patient name — clicking expands it full-width on the same page with a "back to file" control, so the doctor never loses the file context. Patient file: notes full-width on top with add-note at top, findings and anamnesis in two columns below, plus anchor/filter jumps and a prominent panel for critical flags. Client stays mobile-first; after onboarding completes, login routes straight into the AI assistant.
New ground. Chinese LLMs (incl. z.ai) enter the selection — cheap but tested slow; pick a model likely to stay on the market. Self-cost-per-client KPIs come to the admin now; profit-margin alarm added later. Token testing runs via an LLM channel-manager platform with the founders as super-users; a dedicated pricing step follows. First ~12 customers get a free transition month (tentative).
Three customers were surveyed. All three engage intensively by time and generate a lot; some pull the bot into off-topic personal work (website texts, archetype content). The pattern matters for pricing — intensive use drives cost, and topic drift argues either for token / time caps or for structural restrictions.
Off-topic pull vs. cost
Off-topic intensive use is real (website copy, archetype content). If common, it drives cost and argues either for going back to token / time limits, or for structural topic restrictions. Idea raised: an archetype / deep-dive add-on in pricing — pay a fee to go deeper for e.g. a month or 14 days.
Trade-off: recurring "pay again" prompts (like free-tier limits on ChatGPT / Claude) risk annoying users and causing churn. Preference: a few named tiers (intensive / normal / medical) over short 14-day add-ons.
To get real cost data, token testing runs via an LLM channel-manager platform (pre-built API connections) with the founders as super-users working intensively — e.g. 45 min to 2 h per day. Token output converts to CHF; pricing then follows in a dedicated step with a weighting across user types and themes.
Token testing via channel-manager. Founders as super-users, intensive real-world usage (45 min – 2 h / day). Output measured in tokens → converted to CHF.
Weighting by user type. Pricing model uses a weighting across user types and themes plus some math to arrive at a first tier structure.
Free transition month. First ~12 current customers get one free month to bridge the switch and adjust to the new bot — tentative client decision, hedge against churn.
Trial length TBD after tokens. 2 weeks / 1 / 3 months — decide after real cost data lands. A full free year for 12 users is likely not sensible.
Two new signals for the LLM shortlist: cheap Chinese models like z.ai / chat.z.ai are worth testing, and for both US and Chinese models the guiding principle is long-term stability — prioritise something likely to remain on the market so the platform is not dependent on a model that disappears in a year.
Chinese LLMs added
Long-term market stability
The super-admin surface starts filling in. For V1 assume one LLM in the background; the negative-prompt layer is a toggle vs. a default unrestricted model; data location and audit logs are surfaced explicitly — important on the road to medically-certified software and for legal protection down the line.
Super-admin settings
KPI view — self-cost first, revenue later
Data-protection line: super-admin and admin see no findings — only audit-log entries. The doctor sees the timestamped consent confirmation and may then view findings. The consent audit trail is the visible artefact for admins.
Certification is not pursued now. Mendelio's Austrian startup was architecturally prepared for the key requirements, but actually certifying takes about 2–4 months with two staff plus annual audits — a high internal cost, especially at the start. Like ISO, both content documentation and technical implementation must match, and the annual re-certification is detailed. Revisit when the platform moves more toward doctors and away from lifestyle — therapist resellers may not require it.
Not now. The workload and ongoing audit burden don't fit the current stage. Defer until the doctor-facing use case is dominant.
Architecture is ready. Mendelio's stack already covers most of the technical requirements — turning on certification is a documentation + audit exercise more than a rebuild.
SwissDnaCode currently uses only Memberspot's course / unlock module. Options considered: an API integration to Memberspot (little effort, includes payment infrastructure) vs. a simple manual upload in the new software (categories + headline, ~2–3 days effort, fully self-managed). Decision: keep Memberspot at the start — optionally just link to it. The full in-platform academy is Vision.
MVP — link to Memberspot
Vision — full academy
The doctor navigation stays lean: Patients + Settings. Everything patient-related lives inside the patient file — no separate sub-pages, so the doctor never loses their place in the hierarchy. This session locks the specific placement of the doctor bot and the layout of the patient file.
Doctor-bot quick access under the patient name. A field / icon appears directly under the patient's name (e.g. "Maria Schmid"). Clicking it expands the bot to full width on the same page.
Back-to-file control. When the bot is expanded, a "back to patient file" control collapses it. The bot can also expand to a full-size window and collapse back — the doctor stays in the same URL and the same hierarchy the whole time.
Notes at the top, full-width. Notes take the top of the patient file with the add-note field at the top (not the bottom). Timestamped, doctor-editable. Bot answers are copied manually into notes — no automatic partial-answer filing.
Two-column below. Findings and anamnesis sit side-by-side under the notes. Optional anchor / filter buttons (all findings, anamnesis, medication) let the doctor jump within the file.
Critical-flags panel. A prominent field for critical flags (allergies, vaccinations) — balanced against not creating too many panels. Start with the minimum and refine from real practice.
Patient-file contents
Findings · master data · anamnesis / onboarding · notes (doctor-editable, timestamped) · important SNPs · medications and active ingredients. Bot answers are transferred by the doctor into notes — the platform doesn't auto-clip partial answers into the file.
Guiding philosophy: start with the minimum and refine from real practice. Important needs surface in daily use — over-designing panels up front creates clutter.
Mobile-first for the client — this is important. Menu: Start · First Steps · AI Assistant · Findings · Settings, with progress shown (e.g., 2 of 4 complete). Once setup is complete, login routes directly to the assistant. First Steps includes the initial consent to share with Dr. Farkas ("I agree" / "no, later"); the "no, later" path ties into the restriction flow already agreed in Workshop 3. General data-protection / platform rules are accepted earlier, at initial registration.
First-Steps flow
Categorised history & "generate summary"
Categorisation risk: single chats mix topics (sleep + movement + stress). If categorisation only captures half and causes disputes, it's better dropped. Easier in the doctor bot (save as notes); treat it as a training matter, test it, remove if it doesn't hold up. Technically, the AI can detect multiple topics and ask which to address first, then work through them in sequence.
The base design is liked — light, friendly background (keep it light even though the marketing site is dark), the boxes and font are fine. But it reads as typical AI / Claude-generated output: the italic "Welcome back" heading and long em-dashes are tells. Concern that customers might think the platform was thrown together with AI.
Keep the AI design in V1. Building the prototype with Claude is cheaper for the client. The finishing pass happens once the content is right and the logo / colours arrive.
Final product is manual code. Not a 1:1 copy of the AI output — em-dashes / fonts get replaced, browser compatibility and fallbacks are handled properly.
Brand colours & logo. Add brand recognition. The app need not match the corporate identity 1:1 — subtle recognition beats a pixel-copy of the marketing site.
Not urgent. Customers won't see the app until after the intensive testing phase. The finishing pass can run in parallel later.
Nadine delivered the onboarding-workflow document — user flows with master data (mandatory fields), archetype base data (partly optional), medical mandatory fields, consent / data protection, and the dashboard-unlock step (Vision). Categories live inside the chatbot, not as a Vision-only dynamic personal generation on the left side. Optional deepening questions are voluntary and recommended, not required.
Question-design principle
Document sections
Dante delivered a bot-behaviour document with a traffic-light system (levels 1–3): answer behaviour, individual health questions, and edge cases / medical emergencies. Built by iterating a prompt across ChatGPT, Claude and Gemini. Considered a strong, detailed base and a good IT checklist. Emergencies escalate beyond just contacting Dr. Farkas.
A good IT checklist
The document is treated as a strong operational base for the negative-prompt layer — good enough to feed directly into the LLM configuration once the model is chosen. The traffic-light framing keeps the rules readable for non-technical stakeholders while staying precise enough for the prompt team.
Voice is possible but risky due to dialect — Swiss High German is near-dialect and hurts recognition quality. Approach: no dictation for onboarding / anamnesis in V1 (writing is fine there); the customer bot's short questions get a voice option. The doctor bot gets dictation for internal testing first. Recommend High German for quality and doing first onboarding on a PC.
Customer bot — voice on. Short conversational questions tolerate dictation better than long form fields.
Onboarding & anamnesis — voice off. Long, structured input; dialect degrades quality; writing is more reliable.
Doctor bot — internal testing. Dictation implemented in the doctor bot first, tested internally before customer-facing rollout.
Chris to confirm. Feasibility check with the team — dictation quality across dialect zones, error handling, correction flows.
Effort is focused on V1 — Vision stays documented and evolves step by step. The team incorporates both client documents into the prototype and pushes toward a ~80–90% first-version base. Token testing runs in parallel and gates the pricing step.
SwissDnaCode — to-dos
Mendelio — to-dos
Open points — still to decide
The team incorporates both client documents into the prototype; SwissDnaCode clicks through and gives feedback, aiming for a content-correct base at ~80–90% done. Chris uses Monday for an internal meeting — aligning on goals, working toward the offer, and getting into LLM testing quickly.
Monday · internal. Chris uses the Monday slot for a Mendelio-internal meeting to align on goals, work toward the offer, and get into LLM testing.
Mid-week · prototype push. Both client documents incorporated · doctor-bot quick access · two-column patient file · notes-on-top · admin self-cost KPI.
Design in parallel. Not urgent — customers won't see the app until after the intensive testing phase. Finishing pass proceeds later once the logo and colours land.
Next call · Saturday the 11th. Moved from Friday because Nadine has a clinic day. Effort focused on V1; Vision stays documented and will evolve step by step.