Topics discussed:

• Max Nussbaumer showed a patient scribe inside a patient portal: the patient enters or scans data that arrived on paper, it lands as unverified, and a practitioner confirms it later — which is the whole point, since the case he's building for is records that reach a patient in anything but FHIR. Everything is validated against profiles by his TypeScript library that converts StructureDefinitions, and the portal handles DICOM upload with a viewer, document scanning with extraction, and time-limited read-only sharing links. He flagged consent as a bigger topic than a share link and worth its own session.
• Eugene Vestel's Health Claw is a set of skills and agent templates that put a patient's own data behind an always-on agent they control. The agents are deliberately segmented by domain — the one managing prescriptions is not the one giving fitness advice — and data from SMART apps, wearables, claims, labs and pharmacies passes through a guardrail layer that strips PHI first, so a compromised instance doesn't hand over the whole health record. A curator agent checks data quality and tells the patient when a code is wrong or outdated rather than letting it quietly inform medical advice.
• The story carrying the thesis was a surprise $1,000 bill. Gene had the agent pull his family's claims history, analyse the charges against the health plan's own authorisation guidelines and produce a specific action plan — talk to the provider, talk to the insurer — and it saved the whole amount, which he was quick to say is nothing next to what people go bankrupt over. Arjun Sanyal framed the general problem around a newly published book on prior authorisation, where a seriously ill patient described needing a graduate-level understanding of the healthcare system to navigate her own care: the burden has been shifted onto patients, and an agent aligned to the patient is a way to shift some of it back.
• Gene's argument against the obvious alternative: no patient is going to install an app per problem or wire up MCP servers themselves — they aren't aware enough to, and won't be. What works is one always-on agent reachable over Telegram or WhatsApp that hooks into the growing MCP ecosystem on their behalf, which he thinks makes referral management (currently a week of repeated phone calls) the next thing to fall. He also made the ownership case plainly: HIPAA protects data on the provider's side, so once you download it and hand it to a third-party app, that company can do as it likes — self-hosting is the only real control.
• The recurring spec argument resurfaced when Alex Szilagyi noted BMAD's persona-based roles looked like Gene's agent structure. Gene's approach working solo is to skip heavy upfront specs — you go through that thinking anyway while planning — and instead run /init every session so the project's CLAUDE.md compounds, working in small chunks because the more context you hand over at once the harder it is to hit the target. Brian Kaney's counter is that the living spec is exactly that memory, but shared: it lets him run several repo clones as independent developers on parallel features, and it matters more, not less, when you move between model providers.