A free four-part series for physicians and healthcare professionals on integrating AI into clinical practice, research, and daily workflow. All four sessions are complete: watch the full recording or read the written guide for each.
The complete AI as Your Clinical Companion series is four free sessions, roughly three hours total, recorded and written up in full. Session 1 covers AI scribes for clinical documentation, session 2 covers AI agents for email, calendar, and research, session 3 compares Custom GPTs with Claude Skills, and session 4 shows you how to build your own clinical tools by describing them in plain English. Every session has a full recording and a detailed written guide. Not sponsored, not affiliated, built by a practicing oncologist.
Published July 3, 2026
A live walkthrough of how practicing oncologists use AI scribes (Heidi, Open Evidence, Doximity) to automate clinical documentation. Learn about templates, context, data privacy, and real-world EMR integration.
When I transitioned from residency to fellowship, the documentation burden shocked me, and the measured numbers are worse than the felt ones. In the Sinsky time and motion study in Annals of Internal Medicine, physicians spent 49.2 percent of the office day on EHR and desk work versus 27.0 percent on direct clinical face time with patients, plus one to two more hours of data entry at home each night.
Epic event log analysis by Arndt and colleagues in Annals of Family Medicine put it in raw hours: 5.9 hours of an 11.4-hour workday inside the EHR, with clerical and administrative tasks accounting for 44.2 percent of that time. Much of the volume is not clinically necessary either. US clinical notes run roughly four times longer than those written by physicians in other countries. Oncologists, emergency physicians, and hospitalists carry the heaviest burden.
AI scribes are intelligent software systems that automatically transcribe, structure, and populate clinical documentation from physician-patient interactions. Unlike traditional human scribes who manually type notes, AI scribes use natural language processing and machine learning to convert spoken conversations into formatted clinical notes that integrate directly with your EMR.
Three leading platforms, Heidi, Open Evidence, and Doximity Live, offer distinct approaches. Heidi focuses on real-time note generation during patient encounters, Open Evidence emphasizes evidence-based clinical documentation with integrated research, and Doximity Live provides a social-integrated solution.
In my oncology practice: Heidi reduces encounter documentation from 25 minutes to 5 minutes post-visit. The system correctly captures TNM staging, molecular findings, prior treatment regimens, and clinical decision logic. Complex oncology notes that historically required 20-30 minutes now require 3-5 minutes of review and editing.
The real benefit: reclaimed time for clinical thinking, family time, and research activities. The documentation tool becomes transparent, you stop thinking about it and simply benefit from it.
Published July 9, 2026
A live walkthrough of how to use AI agents (Claude Cowork) with Notion, Gmail, Google Calendar, and PubMed to automate email triage, calendar management, and research workflows. Reclaim 8-12 hours weekly.
On top of the record itself sits a second layer. My own week as a fellow ran 6 to 8 hours on administrative tasks that don't involve direct patient care. That's a full workday every week. For a busy fellow, that breaks down to 2 to 3 hours sorting and responding to email, 1 to 2 hours managing calendar requests and coordinating meetings, and 1 to 2 hours tracking research articles, protocols, and reference materials.
Multiply that across a year and you're looking at 312 to 416 hours, the equivalent of 39 to 52 full workdays, spent on tasks that don't require your clinical judgment, your expertise, or your presence. Imagine if you had 7 to 8 extra weeks of time back each year. What would you actually do with that time? More research? Deeper patient time? A life outside medicine?
An AI agent can read incoming emails and categorize them as urgent versus informational versus can-wait. It can draft responses to routine messages, extract key information like lab results and meeting times, update your calendar based on requests, surface relevant research papers related to your patients or projects, and organize everything into a single source of truth.
The magic isn't in doing one thing perfectly. It's in chaining these actions together so you only touch something once, and only when it actually needs your brain.
Before automation: My morning ritual used to consume 45 to 60 minutes scrolling through email, flagging important items, drafting responses, checking calendar conflicts, and searching PubMed for patient cases.
After automation: Claude reviews my email from the previous 24 hours, categorizes messages, drafts responses to routine items, flags urgent items, extracts key information, checks calendar for conflicts, and searches PubMed for relevant articles. I check a single Notion dashboard every morning. Everything that needs my attention is there. Everything else is handled. Time saved: 30 to 45 minutes per day with no context switching.
Published July 16, 2026
A live comparison of Custom GPTs vs Claude Skills. Learn when to use each tool, build both from scratch (no code required), and explore real healthcare examples including clinical trials matching.
Custom GPTs are best for interactive, single-person use. You build a chatbot that lives in ChatGPT's interface. It works brilliantly when one person needs to interact with a focused tool (like patient education, quick decision support, or documentation assistance). The interface is familiar, the learning curve is zero.
Claude Skills are the right choice when you're trying to automate a repeatable process or integrate AI into your organizational infrastructure. A Skill runs unattended, scales to thousands of patients, and integrates directly into your hospital systems via API. When you have 100 patients needing evaluation, not 10, the difference between interactive and automated becomes everything.
Custom GPT Example: One of our fellows created a GPT that explains cancer immunotherapy to newly diagnosed patients. It's embedded in our patient portal. Patients find it intuitive, it feels like ChatGPT, which most have used. It accesses our clinic's protocols and provides consistent, accurate information. The GPT has reduced our clinic's phone time for routine education by roughly 15%.
Claude Skill Example: Our institution runs 30 active oncology protocols. Historically, a research coordinator manually reviewed each patient at each visit: do they meet inclusion criteria? Are they within the age range? Do they have adequate organ function? Do their genetics qualify?
With a Claude Skill, we created a workflow that receives patient data from our EHR, evaluates each patient against all 30 protocols simultaneously, generates a ranked list of matches with reasoning, logs everything with full audit trail, and runs every morning auto-populating a dashboard for our research team. A Custom GPT couldn't handle this. It would require a person sitting at a computer asking ChatGPT about each patient. A Skill runs unattended and scales.
Use Custom GPTs when: One person interacts at a time, the task is bounded and specific, you want a familiar interface, and user feedback is part of the workflow.
Use Claude Skills when: The task repeats constantly, you have many subjects (patients, documents, records), it runs automatically without human intervention, or it needs to integrate with your existing systems.
Published July 25, 2026
The final session: going from user to builder. Describe the tool you wish existed in plain English and watch four platforms build it. A live tour and demos of Lovable, Base44, Replit, and VS Code with Claude Code, plus the privacy rules that apply.
Every clinic runs on duct tape: the risk score you re-derive on paper each week, the equipment sign-out "system" that is a binder by the door, the call rota that lives in one person's head, the spreadsheet that has slowly become a load-bearing monster. You have had the idea for the proper tool for years. It died in a budget cycle, or in an 80 percent fit from a vendor, or in a spreadsheet that grew tentacles.
The friction is worth solving because that is where clinician time actually goes. In the Sinsky time and motion study in Annals of Internal Medicine, physicians spent 49.2 percent of the office day on EHR and desk work versus 27.0 percent on direct patient face time, plus one to two more hours of data entry at home each night.
Vibe coding means describing the software you want in plain English while an AI writes the code you never read. Think attending and resident: you do not write the orders yourself, you describe what you want, review what comes back, and correct it. Describe, look, correct, repeat. The term was coined by AI researcher Andrej Karpathy in February 2025 and was named Collins Dictionary's Word of the Year for 2025.
You need five words of vocabulary and no more. The frontend is your waiting room, the backend is your back office, the database is your records room, an API is the referral fax line between two clinics, and hosting is the building and its street address. Every platform in this session builds all five for you.
Lovable is the design studio: prettiest results, fastest wow, best for calculators and patient-facing pages. Base44 is the turnkey builder with logins and a database included, best for internal team tools (Wix acquired it for roughly 80 million dollars in June 2025, six months after launch). Replit is the full workshop: agent, code, database, and hosting in one tab. VS Code with Claude Code is the pro's bench: steepest curve, no ceiling. Every slide deck in this series was built there.
No PHI, ever, in the builder or the app or the test data. These platforms are not HIPAA or PIPEDA covered by default. A prototype is not a medical device: build for administration and education first. Verify before you share, because AI-built is not correct by default. Veracode's 2025 GenAI Code Security Report found that 45 percent of AI-generated code introduced an OWASP Top 10 vulnerability across more than 100 models tested. And watch the credit meter.