aactingintern·prompting
About this curriculum / Credits & replication / ≈ 3 minutes
06

About this curriculum.

Stanford Internal Medicine AI Curriculum — built by clinicians, refined by use.

On this page
  • · Origin
  • · Course directors
  • · Build this at your program
  • · Citations
  • · Get in touch
§ 1 · Origin

A needs
assessment.

In June 2025, a needs assessment of 34 Stanford IM residents found that nearly all had used at least one AI tool clinically, but mean self-rated AI knowledge was 2.6 of 5. Access without competence is a patient-safety risk.

This curriculum was built to close that gap with five hands-on sessions covering AI foundations, LLM mechanics, QI applications, vibe-coding for research, and red-teaming. The error-finding exercise on this site is the centerpiece of session 3 — LLMs for QI.

34
Stanford IM residents surveyed in the June 2025 needs assessment.
91%
Had used at least one AI tool clinically before any formal training.
2.6 / 5
Mean self-rated AI knowledge — the gap this curriculum was built to close.
§ 2 · Course directors

Built by
clinicians.

Four course directors led design, integration, and clinical framing. Each session was reviewed by practicing hospitalists before it reached residents — and rewritten after each cohort against what actually held up at the bedside.

PH
Poonam Hosamani, MD
Curriculum Design & Medical Education
Hospitalist; led overall curriculum architecture and the cross-session arc from foundations through red-teaming.
KK
Kevin Keet, MD
AI Integration & Curriculum Development
Designed the LLMs-for-QI session and the error-finding exercise; built and maintains this site.
JH
Jason Hom, MD
Clinical Reasoning & Hospital Medicine
Authored the H&P cases and the answer keys; sets the bar for what counts as a real catch versus a near-miss.
NP
Nikhil Patel, MD
Resident Lead, Needs Assessment
Ran the 2025 resident survey that anchored the curriculum; co-designed session evaluation instruments.
§ 3 · Replicate it

Build this at
your program.

You don't need Stanford's resources to run this. Three steps cover most programs — and the order matters less than starting at all.

01
Start with one session.
Pick the format that fits your schedule — you don't need all five to begin. The QI session is the highest-leverage standalone; foundations works as a noon conference.
02
Use real cases.
Modify real H&Ps with deliberate errors. Authenticity drives engagement and clinical relevance — fabricated vignettes don't carry the same weight at the bedside.
03
Provide LLM access.
Free tools work — ChatGPT, Claude, duck.ai. Or run your own custom site like this one. Source code is on GitHub; fork it, swap the cases for yours, deploy.
§ 4 · Citations

References.

Sources behind the needs assessment, the clinical guidelines used to score the exercise, and the broader literature this curriculum draws on.

01
Patel et al. Stanford Internal Medicine AI needs assessment. 2025.
Internal · 2025
02
AHA / ASA guidelines for the early management of patients with acute ischemic stroke. 2019.
Guideline · 2019
03
AHA / ASA guideline for the prevention of stroke in patients with stroke and transient ischemic attack. 2021.
Guideline · 2021
04
Mollick. A Guide to Which AI to Use in the Agentic Era. 2026.
Essay · 2026
05
Kung et al. Performance of ChatGPT on USMLE. PLOS Digital Health, 2023.
Journal · 2023
§ 5 · Get in touch

Take the
materials.

Curriculum materials — slides, facilitator guides, H&P cases with answer keys, and the source code for this site — are available on request. We'd rather see this used at another program than sit in a Stanford folder.

Choosing a Model
End of curriculum
Thank you for reading.
Course home