Josh Mandel, MD

Josh Mandel, MD

@josh.mandel.md

Physician, programmer, and would-be pâtissier/poet. Chief Architect for Microsoft and SMART Health IT. Health tech policy, interoperability, FHIR, AI, web.

993 followers455 following65 posts

... then Claude hit a token output limit while finishing transcribing the solution into a web app. So I spent a third turn saying "Continue," and it emitted the final HTML tags. I'm counting this as a moral success, but it technically violates my pre-specified terms.

New Challenge: "Advent of Claude"! Write a Claude(dot ai) custom style to solve Advent of Code puzzles within Claude's UI. Score: # adventofcode(dot com) stars earned in 2 daily conversation turns.

Thanksgiving prep with LLMs. Here's my recipe: 0. Decide what to cook 1. Put all recipes (or dish ideas) in project context 2. Request * 4 day prep plan * All recipes with consistent formatting * All ingredients in a combined shopping list

View full thread

Prompt 5. Now prepare a combined shopping list organized by grocery department and all duplicate ingredients combined to create totals, but listing the dishes and subquantities where they're used in parens. use check boxes interact app that is formatted for easy printing.

Thanksgiving prep with LLMs. Here's my recipe: 0. Decide what to cook 1. Put all recipes (or dish ideas) in project context 2. Request * 4 day prep plan * All recipes with consistent formatting * All ingredients in a combined shopping list

View full thread

Prompt 4. Now output an artifact containing each recipe with detailed ingredients including quantities, with prep steps written clearly and with quantities repeated inline. Don't bother to confirm with me first, just use your judgment and output all recipes. Separate with ---

Thanksgiving prep with LLMs. Here's my recipe: 0. Decide what to cook 1. Put all recipes (or dish ideas) in project context 2. Request * 4 day prep plan * All recipes with consistent formatting * All ingredients in a combined shopping list

View full thread

Begin by outputting a recipe sketch for each dish. When a recipe is supplied, you should use it as the basis of your sketch. If no recipe is supplied, begin with the recipe sketch that you output so that we understand the overall process.

Prompt 2. Now output a table where the rows are days and each column represents a different dish, so that we can see what prep steps should happen at which cell. (Notes: Focus Sunday on Turkey alone.)

Thanksgiving prep with LLMs. Here's my recipe: 0. Decide what to cook 1. Put all recipes (or dish ideas) in project context 2. Request * 4 day prep plan * All recipes with consistent formatting * All ingredients in a combined shopping list

Sequential prompts. Prompt 1. Our eventual goal is to create a prep schedule starting Sunday night through mealtime Thursday, emphasizing make ahead where it improves or doesn't detract from quality -- while also consolidating prep steps smartly.

@bsky.app feature request: include a "following back" button inline in the "New user followed you" notification screen. (Right now, new follows are grouped into a drop-down, but as you click each one to review — and follow back — the drop-down group is re-filtered and needs to be re-expanded.)

Healthcare data access vs. privacy tensions aren't zero-sum. Better-designed Health Information Networks can bend the tradeoff curve, enabling convenience AND control. No silver bullets, but real progress is possible. User stories & principles: hackmd.io/@jmandel/hin...

hackmd.ioPrincinpled (Q)HIN Design: Convenience AND Control - HackMDAll stories feature Elena, a 42-year-old marketing executive whose healthcare providers participate in AcmeConnect, a Health Information Network (HIN).

EHI Export insight: >100 individual users (i.e., provider org staff) contributed to some aspect of my EHI in the past 6 years of ambulatory care. This diagram shows only users & tasks mentioned at least 10 times. (Task labels generated by GPT-4, based on Epic's EHI schema.)

Just out: www.nature.com/articles/s41... Health AI integration requires workflow & data access. @SMARTHealthIT experience: it takes products, pragmatics, open standards... and regulations. Better alignment & precision on "jobs to be done" => (much) faster timeframes! @mandl @gotdan

nature.comIntegration of AI in healthcare requires an interoperable digital data ecosystem - Nature MedicineElectronic health information, including from electronic health records, is needed to develop AI tools for health, but the seamless flow of data will require standards and interoperability.

Beyond Split Pea Soup: #ChatGPT 1) assembles order sets on-the-fly from primary source 2) controls my web browser through a focused API 3) applies common sense + quantitative reasoning. A strained analogy between online grocery shopping & clinician-facing order entry. youtu.be/447C-f9MJw8

youtu.beAutomating EHR Order Entry, or Shopping for Split Pea Soup with ChatGPTBeyond Split Pea Soup: #ChatGPT 1) assembles order sets on-the-fly from a primary source; 2) controls my local browser through a focused API; 3) applies sens...

It was very cool to see how widely supported #JSON-LD is for recipe content. I wrote a 10 line proxy to extract this content for LLMs. For example... Source kingarthurbaking.com/recipes/mult... Extraction ld.fly.dev?url=https://...

Josh Mandel, MD@josh.mandel.md

Recipes: when I find a keeper, I condense & transcribe (in ink) on a 3x5 card. Which made me wonder... can a custom GPT come close? After 30m fiddling with prompts and proxies: yes, but you be the judge! chat.openai.com/g/g-bdnABvG9... (But you'll have to wrestle the fountain pen from my grip!)

A CMS mandate went into effect in 2021 to ensure that many health plans publish their network details (Provider lists, etc) with a #FHIR API for interoperability. In 2023, things are messy and could use CMS attention. Summary in🧵 Details at hackmd.io/@jmandel/ins...

hackmd.ioInsurance Plan Interop: A Cautionary Tale - HackMDA CMS mandate went into effect in 2021 to ensure that health plans offer #FHIR APIs for their Provider Directory. Let’s explore how things are going.
View full thread

The messy state of FHIR deployments is especially painful given that (in a nearby parallel universe) CMS oversees a *pristine* data resource for insurance plans on the federally-facilitated exchanges, with hyper-specific (non-FHIR) data formats + central aggregation of endpoints.

A CMS mandate went into effect in 2021 to ensure that many health plans publish their network details (Provider lists, etc) with a #FHIR API for interoperability. In 2023, things are messy and could use CMS attention. Summary in🧵 Details at hackmd.io/@jmandel/ins...

hackmd.ioInsurance Plan Interop: A Cautionary Tale - HackMDA CMS mandate went into effect in 2021 to ensure that health plans offer #FHIR APIs for their Provider Directory. Let’s explore how things are going.

Despite the CMS requirement for open, public-access APIs w/ insurance plan details, industry support is inconsistent. When insurers do host APIs, the data are variable and the ability to retrieve full content is limited. It's not clear how CMS intends to enforce rules.

I've been pondering how #Interoperability will change as AI advances. Video: youtu.be/32a7WWMhXd0 Key Points: 1. #chatgpt4 Advanced Analysis Mode is impressive with @HL7 #FHIR data inputs... but 2. Free text can work just as well, using #GPT4 to guide on-the-fly abstraction!

youtu.beAutomating clinical data analysis with on-the-fly data abstractionIn this video, I explore automated capabilities for analyzing clinical data. We begin by running ChatGPT's Advanced Data Analysis tool against a FHIR Bundle,...

For all the good vibes + acceptance here in the skeetiverse, what's with the casual ageism? Scrolling my feed this morning I stumbled over multiple replies tossing age based prejudice into conversations that weren't about age at all. From folks who surely know the ills of discrimination.