Welcome! My name is Harrison Konsker, and I'm an MD Candidate at Stanford School of Medicine, concentrating in Artificial Intelligence and Computational Biomedicine. I work at the HealthRex Lab with Dr. Jonathan Chen, where I build machine learning systems that support real-time clinical decision-making.
Thirteen years ago, I was diagnosed with a rare and aggressive soft tissue tumor in my hand. Ten hospitals across the country insisted that intensive chemotherapy was my only option, until I arrived at Stanford Medicine. There, my care team proposed a revolutionary treatment: high-intensity focused ultrasound, a non-invasive procedure that uses sound waves to destroy tumor tissue. I became the sixth patient ever to receive it. That decision changed the course of my life.
Today, I have returned to Stanford as an MD Candidate, concentrating in Artificial Intelligence and Computational Biomedicine. I work in the HealthRex Lab with Dr. Jonathan Chen, where I build machine learning systems that support real-time clinical decision-making. My current projects separate true urinary tract infections from false alarms in the medical record, predict success at each stage of IVF, and predict antibiotic resistance the moment a patient reaches the emergency department. The thread across this work is the same: using clinical data to support the highest-stakes decisions in medicine. I'm also a VC Fellow at ARTIS Ventures, where I work with the investment team to evaluate early-stage companies in technology, biology, and health. It brings together entrepreneurship, science, and medicine, and I'm eager to learn how research evolves from a paper into a product that reaches patients.
My prior research bridges oncology, genomics, and translational science. I spent a year at the Weizmann Institute of Science investigating low-fidelity DNA polymerases as therapeutic targets in EGFR-mutated lung cancer, and four years at Stanford's Division of Thoracic Surgery with Dr. Joseph Shrager, where my CRISPR/Cas9 work formed the basis of my undergraduate Honors Thesis and a co-first-author manuscript currently under review at Molecular Therapy: Oncology.
Outside the lab, I am the co-founder and president of Fillet for Friends, a national 501(c)(3) addressing food insecurity through wild-caught protein distribution. Since 2014, our team has raised over $550,000 and delivered more than 100,000 meals across four chapters.
Working with Dr. Jonathan H. Chen in the Division of Computational Medicine to build machine learning for women’s health, reproductive medicine, and high-stakes clinical decisions.
Urinary tract infections are the most common bacterial infection in women, yet a positive urine culture is only about 40% specific for a true infection, so a model trained on cultures learns to predict culture positivity, not real disease. I’m building an LLM pipeline that reads clinical notes to separate true UTIs from asymptomatic bacteriuria, with a focus on pregnant and older women, where mislabeling causes the most harm.
An IVF cycle unfolds as a sequence of high-stakes steps, and a patient’s odds shift as it progresses. Working with Stanford’s reproductive endocrinology division, I’m building a dynamic model on the clinic’s live-birth-outcome data that updates its prediction of success at each stage of the cycle, from stimulation through retrieval, fertilization, and transfer, so patients and clinicians can plan with the most current picture at every step.
Patients with suspected infections wait up to four days for culture results, forcing empiric broad-spectrum prescribing and accelerating antimicrobial resistance. We are building ML models that predict the likely organism and its resistance profile at the moment of ED presentation.
Before an AI tool reaches patients, someone has to test whether it truly matches expert judgment. I coordinate a study comparing an AI consult tool’s recommendations against practicing physicians on synthetic clinical cases, measuring where it helps, where it falls short, and whether it changes real decisions.
These are the tools I use to study medicine at Stanford. Three are custom Claude skills I built. Each can be downloaded and used in your own personal Claude account. Four are workflows you can follow with any model. All free to use, modify, and share. Never used AI for studying? This page is built for you. Click any tool below for step-by-step instructions.
Drop in a lecture PDF and your learning objectives, generate a premium-formatted Word study guide (.docx) with MEMORIZE and EXAM TRAP callout boxes, comparison tables, and a Big Picture overview. Saves hours of organizing.
View instructions & download ↓Generate vivid AI image prompts that encode both a medical term's NAME and its MEANING into one memorable scene. Pair with Gemini to render the image. Built for terms that refuse to stick.
View instructions & download ↓Upload a lecture PDF, generate an interactive HTML study guide PLUS a ready-to-import .apkg Anki deck. L2-first Stanford Med workflow: prioritizes integration cards over rote recall. Manually making thousands of cards is becoming extinct.
View instructions & download ↓Turn any lecture into a 15-minute conversational podcast. Interactive mode lets you pause the hosts and ask questions in real time.
View instructions ↓When the lecture doesn't click, generate a visual diagram organized by the most important categories of the content. AI didn't change what's true. It changed how my brain perceives it.
View instructions ↓Turn a lecture's learning objectives into a practice exam that cannot skip anything. A coverage gate catches the plain mechanism and definition objectives an AI would otherwise pass over, which are the ones that show up on exams.
View instructions ↓Interview a patient who has the disease you just studied. You ask the questions, they answer in their own words, and afterward you find out what you never thought to ask. Practice eliciting findings, not just recognizing them.
View instructions ↓Takes a medical lecture (PDF + learning objectives) and outputs a polished, premium-formatted Word study guide (.docx) with Calibri/Arial typography, cardinal-red accents, MEMORIZE and EXAM TRAP callout boxes, comparison tables, and a "Big Picture" overview section. Anti-hallucination by design: every fact comes from your slides or stated learning objectives, never from external board prep.
SKILL.md inside the folder.SKILL.md into the chat, drag your lecture PDF in too, and your learning objectives.~/.claude/skills/ on your computerUse the lecture-to-study-guide skill on this lecture. (with files attached)You give it a medical term (a drug, condition, anatomy term, pathology, microbiology). It returns a vivid AI image prompt that encodes both the term's NAME (via syllable sound-alike hooks) and its MEANING (via scene composition). You then paste that prompt into Gemini to render the actual image. This is how I made the Gordon-Ramsay-in-a-Hunt-outfit image for Ramsay Hunt syndrome.
using nanobanana at the start of your image prompt.
SKILL.md.SKILL.md into the chat.using nanobanana [paste Claude's image prompt here]Every time, with no exceptions:
using nanobanana [Claude's image prompt]The using nanobanana prefix tells Gemini to invoke the Nano Banana image model specifically. Without it, you may get a worse image generator or a text-only response.
The output prompt: Gordon Ramsay in a brown hunting outfit, holding cranial nerve VII like a fishing rod, standing in a forest filled with chickenpox-virus pumpkins. His ear has painful red vesicles. A clock reads "72 hours."
One scene encodes: Ramsay (Gordon Ramsay), Hunt (hunting outfit), CN VII (rod with VII), VZV / chickenpox (pumpkins), painful ear vesicles, 72-hour treatment window.
Upload a medical lecture PDF and generate two things: (1) an interactive HTML study guide for first-pass learning, and (2) a ready-to-import .apkg Anki deck for spaced repetition. The skill uses the L2-first Stanford Med workflow (explained below).
L1 and L2 refer to cognitive levels of a learning objective, not year of medical school:
The skill prioritizes L2 cards (70–85% of every deck) and uses L1 cards sparingly (5–15%) only when a fact is truly irreducible. Why? Because L2 cards force you to retrieve the underlying L1 fact and the relationship that makes it meaningful. You learn the disease, not the slide. The deck stays lean, retention goes up, leech cards go down.
SKILL.md.SKILL.md into the chat. Drag your lecture PDF in too.Turns any lecture into a 10–20 minute audio overview where two AI hosts discuss the content conversationally. Built around Google's free NotebookLM. The killer feature is interactive mode: you can interrupt the hosts mid-conversation with a question and they'll pause, answer, and continue.
Once the podcast is playing:
This moves you from passive listening to active questioning at the exact moment confusion hits.
Takes dense lecture content and generates a visual diagram that maps the concepts spatially. Useful when the lecture has too many moving parts to hold in your head, or when prose isn't sticking.
I'm studying for medical school and the attached lecture covers [TOPIC].
I need a diagram that helps me visualize this.
CRITICAL ANTI-HALLUCINATION RULES:
- Only use facts, terms, drugs, mechanisms, and categories that appear in the attached lecture.
- DO NOT add information from textbooks, board prep, USMLE resources, or your general medical knowledge.
- DO NOT invent or extrapolate relationships, percentages, or values that aren't in the lecture.
- Preserve the lecturer's exact terminology and capitalization.
- If a relationship is unclear from the lecture, leave it out. Do not guess.
- Every label in the diagram must be traceable to a specific slide or sentence in the lecture.
Now generate an image that:
1. Identifies the 3–5 most important categories in the lecture
2. Shows the relationships between them spatially (arrows, hierarchies, branches)
3. Uses color-coding so each category is distinct
4. Includes only the key terms that appear in the lecture
5. Avoids being too dense: readability matters more than completeness
Output: a single clear diagram, not a wall of text. Use a clean medical textbook style.
Before drawing, list which lecture slide(s) each category came from so I can verify.Writes practice questions from your lecture slides with one hard rule: every learning objective gets covered, whether or not it looks interesting. When you ask an AI for practice questions, it gravitates toward the memorable clinical material and quietly skips the plain mechanism and definition objectives. Those are exactly the ones that show up on exams.
Most people ask "write me practice questions about this lecture" and take what they get. The output looks good, so you trust it. The gap only becomes visible on exam day. The fix is not a better prompt, it is a validation gate: generate the questions, then check them against the objectives and refuse the batch if any objective is missing.
Extract every learning objective from this lecture, word for word.
Number them. Do not summarize, merge, or reword them.
If an objective has numbered sub-parts, keep each sub-part.Paraphrasing collapses distinct objectives together, and merged objectives are how coverage gaps start.
For each learning objective, write 4 to 6 practice questions:
at least 3 multiple choice and at least 1 free response.
Scale the count to the content: 4 questions for a single concept on
1-2 slides, 5 for a multi-part objective on 3-5 slides, 6 for a
multi-step or matrix topic on 6 or more slides.
Rules:
- Every question must be answerable from the slides. Cite the slide number.
- Each objective needs at least one recall/understand question AND one
apply/analyze question.
- You may not skip an objective for being "too basic" or "just a definition."
- Every multiple-choice question needs a one-line rationale for each option,
explaining why the wrong ones are wrong.Start a fresh conversation, attach the objectives and the questions, and ask:
Here are the learning objectives and the practice questions.
For each objective, list how many questions cover it.
Flag any objective with fewer than 4 questions or zero free-response
questions. Do not fix anything. Just report the table.Use a new conversation so the model audits the work instead of defending it. Send the gaps back for regeneration, then re-audit until the table is clean.
For each question, quote the exact line or figure from the slides that
supports the correct answer. If you cannot find slide support, mark the
question UNGROUNDED.Anything ungrounded is either invented or pulled from outside knowledge. This is the highest-value check in the workflow, because a plausible wrong question teaches you a plausible wrong fact.
Check for: answers clustering on one letter (spread them); the correct
answer being noticeably longer than the distractors (equalize the lengths);
parenthetical hints that give away the answer (remove them); two options
that mean the same thing.Length bias and letter bias are the two habits that make AI-written questions trivially easy in a way that flatters you during review and fails you on the exam.
Lecture slides may be proprietary to your school, and some clinical material contains protected health information. Confirm with your course director or professor before putting a lecture into an AI tool, and never upload anything with patient identifiers.
Turns a lecture into a patient you can interview. You ask the questions, the patient answers in their own words, and when you say "end encounter" the same model steps out of character and coaches you on what you missed. Six steps can teach you the facts about a disease. None of them teach you to sit across from the person who has it.
A practice question hands you the findings and asks for the diagnosis. A real patient hands you nothing. They tell you their ear hurts. Whether you ever learn about the rash, the facial weakness, or the hearing change depends entirely on what you think to ask. That gap between recognizing information and eliciting it is where most of the actual skill of medicine lives.
This is one prompt, not four. Both roles are loaded at the start, and a keyword moves between them. Attach the lecture, or just name the condition:
Setup: Specialty [cardiology · pulmonology · neuro · OB...], patient age,
difficulty [beginner · intermediate · advanced], scope [history · +exam · +workup].
Role & Goal: You're a realistic standardized patient during the encounter,
and my coach afterward.
Ground Rules: Stay in character and talk like a real patient, not a textbook.
Never volunteer clues, reveal the diagnosis, or invent facts.
The Debrief: When I say "end encounter," step out and coach me: what to keep,
stop, and improve, the high-yield questions I missed, and a brief differential.
Start when I say I'm ready.Two rules carry most of the weight. Never volunteer clues is what leaves you something to practice; without it the model hands you the full case in its first reply. Talk like a real patient, not a textbook is what makes you translate: a patient says their face feels heavy, not that they have unilateral facial paresis. The Setup line is what makes this reusable, since changing four values gives you a different encounter rather than a different disease.
Interview them. Ask open questions first, then narrow. Push on anything vague. You will feel the difference immediately: you have to decide what to ask. If you never ask about hearing, you never find out about the hearing change, and that omission is the lesson. When you are ready, state your diagnosis and your reasoning before asking for feedback. Committing first is what makes the feedback land.
end encounterTwo words. The model drops the patient and becomes your coach: what to keep, what to stop, what to improve, the high-yield questions you missed, and a brief differential.
The trigger is the design, not a convenience. Because both roles are loaded from the start, the switch happens on your command rather than the model's judgment, so it never breaks character early to reassure you. And because you choose when to end it, you have already committed to your own read of the case before any coaching arrives. Say your diagnosis and your reasoning out loud before the keyword.
Confirm with your course director or professor before uploading a lecture, and never upload anything containing patient identifiers. Ask for a generated patient rather than pasting in a real case. The simulated patient is also not a real patient: if a presentation does not hang together clinically, check your slides. This builds the habit of asking systematically, but it does not replace standardized patients, clinical skills coursework, or real patient contact.
A national 501(c)(3) addressing food insecurity through wild-caught protein distribution. I co-founded the organization in 2014 and currently serve as president, leading a 10-person leadership team across four chapters.
Endorsed by the Florida Fish and Wildlife Conservation Commission and U.S. Representative Ted Deutch (Florida 22nd District).
Alongside research, I stay close to patients and community through clinical volunteering, teaching, and student leadership.
I am always open to conversations focused on the intersection of medicine and machine learning, especially with clinicians, researchers, founders, and operators whose work touches clinical decision-making, computational biomedicine, or how we train the next generation of physicians. If that is you, I would love to connect.