This week: the MHRA says a scribe that only transcribes is not a medical device and Healthwatch says patients are the ones catching its errors, a GP builds a telemedicine platform in four evenings, and four new names join the directory.
Top signal
Not a device, says the regulator; patients are catching the errors, says the watchdog. Healthwatch England, the statutory patient champion for the NHS, told the Guardian on 31 August that it has heard "multiple stories from patients who have noticed these errors when a health professional hasn't": a scribe summary that turned an MRI result of "null demyelination" into demyelination, a prescribed drug swapped for one with a similar name, a summary letter that dropped a consultant's instruction to get a repeat migraine prescription from the GP. In every case the patient, not the clinician, caught it; 27 different scribes are already in use in English general practice and hospitals. The regulatory line had been drawn five weeks earlier: on 29 July the MHRA, with NHS England, confirmed that ambient voice technology "intended solely for transcription, summarising of clinical conversations, drafting letters, or suggesting clinical codes for a clinician to review" is not regulated as a medical device in Great Britain, while a product that supports diagnosis, treatment or prevention, or takes automated action such as placing orders without clinician review, is. Healthwatch finds that "worrying". The MHRA's own text puts the weight where it always was: clinicians remain responsible for "reviewing and verifying AI generated transcripts, summaries and other outputs before they are used in patient care", a responsibility that "is unchanged by the guidance". If you are building a scribe, or bolting one onto a workflow: transcription-only keeps you out of device regulation, and every uncaught error goes into the record under your name. The moment your tool suggests a diagnosis, a dose or an order, it is a device. The Guardian
Shipped this week
- My Doctor: GP partner Fahim Hussain MBChB MRCGP (Northern Health, West Yorkshire), who had never written a line of code, built the practice's private telemedicine platform on Replit in four days of evenings after his daughters were asleep: doctor booking, prescription requests, a subscription tier, Fitbit data, nutrition and habit trackers and a medication database with an AI side-effect checker, 16 features in all. Agencies had quoted £75,000 to £100,000 and 12 to 18 months; Replit's customer story puts his total spend at £175 (Forbes, which picked the story up on 26 August, says £165). Two caveats: the account is the vendor's own case study, and it describes the platform as "ready for launch", not live. Replit customer story
- ClinicalSim: paediatric intensivist and palliative care physician Lauren Rissman, who nearly died after her daughter's preterm delivery and credits her husband's hours of unheard advocacy with her survival, co-founded (he is the CEO) and shipped a voice platform: clinicians and students talk through cases with synthetic AI patients, and every score is backed by their own quoted words against a published framework or the institution's own policy. "Communication isn't a soft skill. It changes outcomes," she posted on Thursday. Over 25 academic medical centres and children's hospitals are piloting it, per the site; no outcomes published yet. X
- mychart-takeout: physician and programmer Josh Mandel published a tool that exports all of your own health data from an Epic MyChart patient portal by replaying MyChart's own internal API calls from inside your authenticated browser tab: it never sees your password, the data stays on your machine, and no language model sits in the export path. Tested against several production Epic instances, pushed 26 August; no licence file yet. GitHub
- And in the workshop, Milan district doctor (medico di distretto) Leonardo Pegollo is stress-testing his clinical workflow app: "a few thousand patients" (test data, we assume; he does not say), simulated three-day usage runs and deliberate outages in different parts of the app. No name or link yet. X
Build safely
- Cleared is not tested: of 1,357 AI-enabled medical devices cleared or approved by the FDA up to 5 December 2025, 34 (2.5%) were linked to a registered prospective trial, 12 (0.9%) had a peer-reviewed publication and 3 (0.2%) had evaluated a patient-centred outcome such as mortality, morbidity or readmission; most of the studies that do exist are observational, small and homogeneous, and often exclude vulnerable groups. Abulibdeh, Cajas Ordóñez, Celi and colleagues in PLOS Digital Health on 19 August, open access. The line to pin above the desk: "Readiness should no longer be defined by FDA clearance alone, but by demonstrated, durable, and equitable benefit to patients." PLOS Digital Health
- Where the MHRA draws the line: the 29 July guidance behind this week's top signal, in the regulator's own words: it "does not change the law", it sets out how existing device regulation applies to ambient voice technology in Great Britain. The sentence that matters most to a builder asks NHS boards to make sure "any change in a product's functionality that may alter its regulatory status is identified and reviewed". The feature you add in week three can change what you are legally shipping. MHRA
- The bill arrives after the build: Micky Tripathi PhD, Mayo Clinic's chief AI implementation officer, told Becker's on 31 August that Mayo has 128 clinical AI solutions in practice, about 70% built in-house ("we are ultimately, for those solutions, the legal manufacturer"), and that "the cost of maintaining these systems is far higher than I think any of us in the industry really thought": the compute to keep models running, the people to watch for data and model drift, the data infrastructure underneath. His engineers "can never let go of the product" because the drift has to be watched. A solo clinician-builder has the same three lines on the bill, at a scale of one. Becker's
Tools & guides
- Vibe coding for ophthalmologists: Fares Antaki, Victor Bellanda, Pearse Keane and Ariel Yuhan Ong, in Current Opinion in Ophthalmology (ahead of print, 3 September), examine vibe coding through one specialty's lens: the value sits in "niche, subspecialty-specific problems that are too small or too specialized to attract commercial development"; for low-risk tools that do not touch clinical decisions, clinicians "may in certain settings be able to deploy vibe-coded tools independently"; for anything higher risk it "does not replace professional software development", but it makes needs concrete and prototypes real. Abstract open, full text paywalled. Current Opinion in Ophthalmology
- Vibe coding for statistical analysis: Dante Anthony Tolentino, Elizabeth Kohout, Paul Boy and Hafifa Siddiq in Nursing Research (18 August, open access) work through a human-in-the-loop workflow for vibe-coded data analysis with Cursor: framing the question, operationalising variables, documenting decisions, retrieval-augmented prompts. A logistic regression on a public teaching dataset reproduced the reference coefficients, p values and fit statistics, and still shipped an error in the predicted-probability confidence interval that the first review missed. That last sentence is the method. Nursing Research
- MedSci Skills: Yoojin Nam MD (radiology, Asan Medical Center, Seoul) maintains an MIT-licensed pack of 59 agent skills for clinical research, from literature search and sample size through de-identification, statistics and figures to 49 reporting-guideline and risk-of-bias checks and journal selection, for Claude Code, Codex, Cursor and Copilot; 282 stars, pushed this week, preprint on arXiv. Built and, per the README, tested on real publications. GitHub
Events
- Sep: OpenClinTech's Clinician Engineer Bootcamp is taking registrations for its September cohort: an entry-level programme teaching clinically trained people to build full-stack applications, with collaborative build exercises and short pre-recorded lessons added after the first cohort; now $99 (the first cohort was free), sign-ups closing in early September. OpenClinTech
- 3 Oct (Antwerp): Care & Code Clinical Build Day: doctors, nurses, pharmacists and physiotherapists vibe-code a working AI care tool in one day. Disclosure: this is our own event. careandcode.be
- 20 Oct (Nashville): the AAFP's FMX AI Hackathon: four hours, no coding experience required, bring a primary-care pain point and leave with a prototype concept; free for FMX attendees, capped at 100 places. AAFP
New in the directory
- Steve Rover, RN (France): home-care nurse in Arcachon with twenty years of hospital work behind him and five iOS apps built solo with Claude Code and Swift, among them Pocket Nurse (eight tools for nurses) and NGAP Infirmier Libéral (home-care billing codes). Featured in Issue 1; now a verified entry. X
- Colin Whaley, MD MSc (Canada): internal medicine resident at the University of Toronto who built MedTools.ca, a self-hosted home for the templates and clinical resources he kept losing when preceptors' documents went offline, plus an AI prednisone-taper calculator and a bronchoscopy diagram maker. colinwhaley.com
- Jonathan Adly, PharmD (USA): pharmacist, founder and CEO of TJM Labs, whose pharmacy AI agents automate intake, refill and prior-authorisation admin; Vaccine Bot, launched 16 July, enters vaccination records into the pharmacy system after pharmacist review. Before that, personal coding projects on GitHub and a blog on learning to code as a pharmacist. TJM Labs release
- Luke Hale (United Kingdom): NHS doctor and designer in London, creator of Lungy, the app that recognises and measures breathing through the phone (Show HN 2023, 666 points; relaunched for anxiety and sleep in 2025), and of Lungy Health for asthma and COPD with UCL. lukehale.com
- The directory now counts 48 clinician-builders with verified, evidence-graded ships: four joined this week and one entry was unpublished at the person's own request, so the net is two more than last week. Browse it, forward this to a colleague who ships, and reply if you know someone who belongs in it. cliniciansthatcode.com
From us
- Which platforms sign a BAA: 17 platforms in the clinician's vibe-coding stack, from Lovable, Bolt.new and Replit through Cursor and Claude Code to Supabase, Vercel, the OpenAI and Anthropic APIs and GitHub, checked against each vendor's own terms for who signs a Business Associate Agreement, at what tier and price, with GDPR Article 28 columns for EU builders. Every row carries the date it was verified, a watcher re-reads the vendor pages three times a week and flags drift, and the table is open data (JSON, CC BY 4.0). We built it because every other answer to "does X sign a BAA" is written by someone selling the compliance layer. cliniciansthatcode.com
- The Library: the full month-by-month archive of everything the monitoring pipeline has surfaced since May 2026, open and ungated; a story covered by two hosts now shows once, with "also covered by". The 28-day Pulse window was throwing the reading away; the Library keeps it. cliniciansthatcode.com