AI Workflows for Clinicians — ObGyn Intelligence

The Three Workflow Modes

Every clinical AI task fits one of three patterns. Matching the task to the right mode determines quality of output.

Focused Session

A fresh conversation, no setup, no memory. Like texting a consultant who knows everything but nothing about you specifically.

Best for: DDx brainstorm · terminology · quick literature check · drafting a patient message · explaining a lab value

ObGyn example: "What are causes of elevated AST in the third trimester? Rank by probability."

Think of it as: A smart consultant you text once

Context Workspace

Upload files — protocols, guidelines, spreadsheets, drafts — and Claude works within that context. The more you give it, the better it performs.

Best for: Reviewing a guideline PDF · editing a manuscript · analyzing a dataset · comparing two protocols

ObGyn example: "I uploaded our preeclampsia protocol and the 2024 ACOG Practice Bulletin. Flag any discrepancies."

Think of it as: An assistant who has read your folder

Long-Term Projects

A persistent workspace that remembers your instructions, files, and style across every session. Set it up once; it works for you indefinitely.

Best for: AJOG reviewer workflows · research programs · teaching material libraries · ongoing writing projects

ObGyn example: "My AJOG reviewer project includes my criteria, past reviews, and formatting preferences — ready on every open."

Think of it as: A permanent digital team member

The key insight: AI output quality depends less on which model you use and more on how much clinical context you provide. Focused Session gives minimal context. Projects give maximum context. Match the mode to the task.

What Each Mode Cannot Do

Mode	Does NOT have	Workaround
Focused Session	Memory of past sessions, your files, your style	Paste relevant context at the top of each message
Workspace	Persistence across sessions (files reset)	Re-upload files each session or use Projects
Projects	Real-time web access (unless Search is enabled)	Enable Web Search tool, or paste current data in

Privacy reminder: Never share PHI in any AI conversation — not even in Projects. Use de-identified or synthetic examples for teaching and case work.

Claude-Specific Features Worth Knowing

Most clinicians use 20% of what Claude can do. These features change the ceiling.

🧠

Extended Thinking

Claude reasons through complex problems step by step before answering. Surfaces contradictions, weights evidence, flags uncertainty. Turn it on for diagnostic reasoning, guideline critique, or multi-variable decisions.

Claude.ai Pro

🔍

Web Search

Claude searches the current web in real time. Use for recent trial results, drug updates, FDA alerts, or anything post-2024. Toggle it on — it's off by default in most conversations.

Available free

📄

Document Upload

Upload PDFs, Word docs, spreadsheets, images. Claude reads them in full. Upload a guideline and ask it to compare against your protocol. Upload a dataset and ask for statistical analysis.

Available free

🖥️

Artifacts

Claude generates a separate panel for documents, code, HTML tools, spreadsheets, and presentations — not just chat text. Ask it to build a patient handout, a risk calculator, or a slide deck and it appears in a live preview.

Available free

📁

Projects

A persistent workspace. Upload your files, write instructions, and Claude carries everything across every session. Ideal for residency teaching libraries, manuscript projects, or recurring review workflows.

Claude.ai Pro

🧬

Memory

Claude remembers facts about you across separate conversations — your specialty, your writing style, your recurring collaborators. It builds a profile over time. You can view and edit it anytime.

Claude.ai Pro

📊

Deep Research

Multi-step research that searches dozens of sources, synthesizes, and produces a formatted report with citations. Takes 5–30 minutes. Ideal for systematic literature reviews, guideline backgrounds, or grant prep.

Claude.ai Pro

💻

Claude Code

A separate command-line tool that writes, runs, and debugs code. Useful for data analysis pipelines, CDC natality database queries, statistical scripts, or building custom clinical tools. Separate from the chat interface.

CLI Tool

Which model to use? Claude Sonnet 4.6 handles most clinical tasks well. For complex reasoning, manuscript critique, or multi-document analysis, use Claude Opus 4.6. Extended Thinking amplifies both. For quick, high-volume tasks, Claude Haiku 4.5 is faster and cheaper via API.

Plan Levels: What You Get at Each Tier

Claude is available at four levels. The free tier is a genuine starting point. Pro is where clinical workflows become sustainable. Team and Enterprise add organizational controls.

Feature	Free	Pro ~$20/mo	Team ~$30/user	Enterprise Custom
Conversation access	✓ Limited daily	✓ Higher limits	✓ Higher limits	✓ Custom limits
Document upload	✓	✓	✓	✓
Web Search	✓	✓	✓	✓
Artifacts	✓	✓	✓	✓
Projects	—	✓	✓	✓
Memory	—	✓	✓	✓
Extended Thinking	—	✓	✓	✓
Deep Research	—	✓	✓	✓
Claude Opus access	—	✓	✓	✓
Shared Projects (team)	—	—	✓	✓
Admin controls & analytics	—	—	✓	✓
SSO / SAML	—	—	—	✓
Data privacy guarantees	—	—	✓	✓
Custom retention policies	—	—	—	✓

For clinical use: Pro is the minimum useful tier. Projects and Memory are what make Claude work as a clinical tool rather than a general chatbot. If you share workflows with a department or team, Team adds collaboration without requiring everyone to manage their own setup.

Settings: Do These Before Your First Clinical Session

Most clinicians skip settings entirely and wonder why Claude does not seem to know anything about them. The settings are where you stop being a stranger to your own AI tool. These take under 10 minutes and change every subsequent interaction.

👤

Settings → Profile (User Preferences)

Do first

This is Claude's permanent briefing about who you are. It loads before every single conversation — including ones with no other context. Without it, Claude treats you as a generic user.

Write in your profile:

Your specialty, subspecialty, and clinical role ("I am a practicing obstetrician and maternal-fetal medicine specialist at an academic medical center")
Communication register ("Respond at a peer-to-peer level. Full medical terminology. No explanatory asides for concepts I already know.")
Default citation format ("Vancouver format for all references. Flag any unverifiable DOI with ⚠️ UNVERIFIED.")
Format preferences ("Default to prose. Tables for comparative data. Avoid bullet lists unless I ask.")
What to suppress ("Omit generic disclaimers about consulting a physician — I am the physician.")

🧠

Settings → Memory

Review regularly

Claude builds a memory of facts about you from your conversations — collaborator names, platform preferences, ongoing projects, recurring rules. You can view every memory entry, correct errors, and delete anything outdated or wrong.

Open Settings → Memory and read every entry — errors accumulate silently
Delete anything outdated, incorrect, or that you do not want Claude to act on
Add key facts Claude has not yet learned by typing them directly into a conversation ("Remember that my journal is called X")
Critical rules should also live in Project Instructions — Memory can lag by days after a conversation

🔧

Settings → Features

Configure once

Controls which tools and capabilities are active by default. Most are off until you turn them on — and they stay off even when they would have helped.

Artifacts — turn on. Enables Claude to generate live documents, calculators, and tools in a side panel instead of plain text
Web Search — turn on. Claude will search when relevant; you can always ask it not to for a specific query
Memory — turn on if you want Claude to build a persistent profile about you over time
Extended Thinking — available per-conversation; does not need to be set globally, but know where the toggle is

✍️

Settings → Style

Optional but powerful

Upload a sample of your own writing and Claude calibrates its output to match your voice and register. Most useful for anyone using Claude to produce content — newsletters, manuscripts, patient communications, LinkedIn posts.

Paste 2–3 paragraphs of text you have written that represent the voice you want Claude to match
Claude analyzes sentence structure, terminology, tone, and register — then uses it as a style anchor
Particularly useful when the output has your name on it

The 10-minute rule: Profile + Memory check + Features toggle = under 10 minutes, done once. A clinician who spends 10 minutes on Settings gets a fundamentally different Claude from one who never opens them. This is the highest-ROI action in this entire guide.

Projects

A Project is a persistent workspace inside Claude that carries your instructions, files, and context across every session — indefinitely. Without a Project, Claude starts fresh each time you open a new conversation. With one, it already knows your role, your standards, your files, and your workflow before you type a word.

The core difference: A regular conversation is a single session — context disappears when you close it. A Project is a permanent workspace — everything you put in it is available every time you return.

What a Project contains

📝

Project Instructions

A text field where you write your standing rules. Who you are. What you need. How Claude should behave. What it should never do. This loads into every conversation in the Project automatically.

📁

Uploaded Files

PDFs, Word docs, spreadsheets — any reference material Claude should always have access to. Guidelines, protocols, past work, templates. Upload once; available in every session inside this Project.

🔁

Conversation History

All conversations inside a Project are grouped together. You can return to any prior conversation, continue where you left off, or start a new one — the instructions and files are always there.

⚙️

Skills (named behaviors)

Named instruction blocks you define inside the Project Instructions. Each has a trigger word. When you type the trigger, Claude executes the behavior you defined for it. See the Skills section below.

How to create a Project

Open Projects

On claude.ai, find "Projects" in the left sidebar. Click "New Project." Give it a descriptive name tied to a specific workflow — not "My Project" but "Journal Peer Review" or "Resident Teaching Cases."

Write your Project Instructions

This is the most important step. Start with: who you are, what this Project is for, how Claude should behave, and any non-negotiable rules (citation format, what to flag, what to never do). Be specific. Two paragraphs beats one vague sentence.

Upload your reference files

Any document Claude should always have: guidelines, protocols, past work examples, data dictionaries, templates. These are available in every conversation inside this Project — you never need to re-upload them.

Start a conversation inside the Project

Click "New Chat" inside your Project. Claude now has everything: your instructions, your files, your rules. Every conversation you start here begins from that foundation — no re-explaining required.

→

Maintain it

Revisit your Project Instructions every few weeks. Update files when guidelines change. Add examples of output you liked. A Project that improves over time is fundamentally different from one you set up once and forget.

What to put in Project Instructions

Think of Project Instructions as the briefing you give a highly capable colleague before they join your team. Cover these five things:

Element	What to write	Example
Your identity	Who you are and what this Project is for	"I am a practicing obstetrician at an academic medical center. This Project supports my monthly journal peer review work."
Behavioral rules	How Claude should communicate and at what depth	"Respond at a peer-to-peer clinical level. Use correct medical terminology. Do not explain foundational concepts unless I ask."
Output standards	Format, citations, length, structure	"All references in Vancouver format. Flag any unverifiable DOI with ⚠️ UNVERIFIED. Default to prose, not bullet lists."
Hard constraints	What Claude must never do in this context	"Do not fabricate citations. Do not add generic medical disclaimers. Do not hedge when evidence is clear."
Skills (optional)	Named behavior blocks with trigger words	"When I write 'ReviewStart:', treat the uploaded file as a manuscript and begin a structured peer review using these criteria: [criteria]"

Skills: Named Behaviors Inside a Project

A Skill is a named instruction block that lives in your Project Instructions and activates when you type its trigger word. It is not a special Claude feature — it is a convention you create. The power is in the consistency: the same task, executed the same way, every time, without re-engineering the context.

How it works: In your Project Instructions, write a block that starts with a trigger word and defines exactly what Claude should do when you use it. Then, in any conversation inside that Project, type the trigger word — and Claude executes the behavior you defined. One word replaces a paragraph of instructions.

Without a Skill — what you type every time

Please summarize this paper for my newsletter. Write it at a 7th grade reading level, about 400 words, in three sections: a summary, what it means for patients, and my take. No jargon. Vancouver citation for the paper itself.

With a Skill — what you type every time

Summ [upload PDF]

To create that Skill, you add this block once to your Project Instructions:

SKILL: Summ When I type "Summ" and upload a PDF, summarize the paper for my newsletter. Format: 400–500 words total at 7th–8th grade reading level. Three sections: Summary (150 words), What It Means (150 words), My Take (150 words). No jargon. One Vancouver citation for the paper. Deliver as structured prose, no bullet lists.

The rule: One Skill per recurring task type. The moment you notice yourself typing the same instructions twice, that is a Skill waiting to be written. Build one per week until retyping context becomes a thing of the past.

How to Create, Save, and Share a Skill

A Skill is plain text. You write it once, paste it into your Project Instructions, and Claude reads it every session. No special files, no technical setup. Here is exactly how.

Step 1 — Write the Skill block

Use this exact structure. The trigger word is what you will type in any conversation to activate it.

SKILL: [TriggerWord] When I type "[TriggerWord]", do the following: [Your complete instructions here — be specific. Include: output format, length, reading level, citation rules, what to include, what to omit, and any non-negotiable constraints.]

Step 2 — Paste into Project Instructions

Claude.ai → Projects → your Project → click the pencil/edit icon on Project Instructions → paste the Skill block at the bottom → save. That is it. Claude reads it at the start of every conversation in that Project.

Step 3 — Test it

Open a new conversation inside the Project. Type only the trigger word and see what Claude does. If it does not behave as expected, go back to Step 1 and make the instructions more specific. One iteration is usually enough.

Saving your Skills outside Claude

Your Skills live inside Claude's Project Instructions — they are not stored as separate files. To keep a backup or share them:

Copy the full Skill block text from Project Instructions
Paste into a plain .txt or .md file on your computer
Name it clearly: LitSum_skill.txt, GuideAudit_skill.txt
Store in a folder: My Claude Skills/
When starting a new Project, paste Skills back in from this folder

Sharing Skills with colleagues

A Skill is plain text — share it like any text document.

Email the .txt file directly
Share via a Google Doc or Dropbox link
Your colleague pastes it into their own Project Instructions
They get the exact same behavior from Claude — consistent across your whole team or department

One Project, multiple Skills: Stack as many Skill blocks as you need in a single Project Instructions field. Each one is activated by its own trigger word. A mature Project might have 5–10 Skills covering every recurring task type in your workflow.

5 Skills Every Clinician Should Build First

These five cover the tasks every clinician repeats most: reading the literature, critiquing guidelines, communicating with patients, teaching, and verifying references. Click "Build This Skill" on any card — Claude will interview you and generate a customized Skill block ready to paste into your Project Instructions.

Your Claude Setup: The Configuration Stack

Most clinicians open Claude and start typing. That is like showing up to a consultation without telling the consultant who you are, what specialty you practice, or what you already know. The configuration stack is how you tell Claude who it is working with — once, permanently, at every level.

The hierarchy: Claude reads your configuration in layers, from broadest to narrowest. Each layer narrows and sharpens the one above it. A fully configured Claude is a categorically different tool from an out-of-the-box one.

Layer by Layer: What to Put Where

Personality and Style: Making Claude Sound Right

Claude's default register is helpful generalist. You can shift it — in Settings, in Project instructions, or at the start of any session — toward the tone and depth your work requires.

🎯

Depth and register

Tell Claude the level you want. "Respond at the level of an MFM colleague — full terminology, no explanatory asides, assume I know the basics" produces a completely different output than the default.

Respond at the level of an experienced MFM specialist. Use correct obstetric and medical terminology. Do not explain foundational concepts unless I ask. If evidence is contested, say so explicitly.

✍️

Writing style

For academic writing, give Claude your voice: a paragraph you have written, a paper you have published, a phrase you use habitually. It calibrates to that style faster than any description of it.

Here is a paragraph from a paper I published. Match this voice and register in any writing you produce for me: [paste paragraph]

🚫

What to suppress

Claude defaults to disclaimers, caveats, and hedges. For clinical expert users, these add noise. You can suppress them explicitly — and should, for your Project instructions.

Omit generic disclaimers about consulting a physician — I am the physician. Do not hedge when evidence is clear. Flag uncertainty when it is real and clinically relevant, not as a default reflex.

📐

Format defaults

Claude defaults to bullet points and headers. Tell it what you actually want — prose, tables, numbered lists, or specific structures like SOAP or PICO. Set this in User Preferences or Project instructions.

Default to prose, not bullets. Use tables when comparing data across multiple variables. Use numbered lists only for sequential steps. Never use headers for responses under 400 words.

Skills: Giving Claude a Standing Job Description

A Skill is a saved instruction set that pre-loads specific behaviors for a recurring task type. Instead of re-engineering the context every time you need a literature summary or a manuscript review, the Skill does it for you. Think of each Skill as a standing referral to a sub-specialist who already knows your standards.

Skills live inside Projects. Each Project can have multiple Skills. A Skill is simply a named block of instructions in your Project instructions field — Claude reads them and knows what each trigger word means. "ObSum" triggers the literature summary protocol. "GuideEval" triggers the guideline audit. The trigger is the context. The Skill is what Claude does with it.

The Full Configuration Checklist

Run this once when you set up Claude for clinical use. Revisit it whenever your needs change.

Context Engineering: The Skill That Determines Everything

The industry used to call it "prompt engineering." That framing was too narrow. What you are actually doing is engineering the context Claude thinks inside. The model is fixed. The context is yours to build.

The foundational principle: Claude has no clinical experience, no memory of your patients, no knowledge of your hospital protocols, no awareness of your subspecialty norms, and no idea what you already know. Every piece of context you omit, it fills in with statistical averages from its training data. Give it your context, and it reasons inside your world. Omit it, and it reasons inside a generic one.

The Seven Layers of Clinical Context

A fully engineered clinical context is built from seven layers. Most clinicians provide two or three. The difference in output quality is not marginal — it is categorical.

Before and After: The Same Question, Two Different Worlds

These paired examples show what changes when context is engineered deliberately. The question is identical. The output quality is not.

The SCOPE Framework for Clinical Context

A five-element checklist you can run mentally before any clinical prompt. If you can answer all five in under 30 seconds, your context is ready.

Setting

Who am I? What is my role, specialty, and clinical context? Academic MFM vs. community ObGyn vs. resident on call — each implies different depth, different defaults.

Clinical picture

The full patient situation: GA, parity, comorbidities, medications, relevant history, current vitals and labs. Not what you think is relevant — all of it. Let Claude decide what matters.

Objective

One precise task. "Help me" is not an objective. "Rank differentials by probability with distinguishing features" is an objective. The more specific the task, the more targeted the response.

Parameters

Constraints and format. Word count, reading level, citation format, what to include and exclude, what to flag. Without parameters, Claude fills in defaults — which may not match your needs.

Example or anchor

When style, tone, or structure matters, give one example of what good looks like. A past review, a handout you liked, a paragraph in your voice. One example is worth a hundred adjectives.

The Six Most Common Context Mistakes

Mistake	What it looks like	The fix
No role assignment	"What causes elevated liver enzymes in pregnancy?"	Start with: "I am an MFM specialist at an academic center. My patient is..." — Claude calibrates depth, terminology, and assumptions accordingly.
Omitting gestational age	"Patient has elevated BP and headache"	GA changes everything in obstetrics. 20 weeks vs. 36 weeks with the same presentation triggers completely different workups and management thresholds.
No output format	"Explain the management of preeclampsia"	Specify: "Format as a decision tree / three-column table / bulleted protocol / 200-word summary for a patient." Format shapes what Claude produces as much as the question does.
Asking for everything at once	"Tell me everything about shoulder dystocia"	One specific task per prompt. "List the HELPERR maneuvers with the evidence grade for each" produces a usable answer. "Tell me everything" produces a textbook chapter.
No citation instruction	Accepting numbered references without specifying format or verification	Always add: "Vancouver format. Flag any PMID or DOI you cannot verify with ⚠️ UNVERIFIED." AI models can and do fabricate plausible-looking citations.
Treating uncertainty as failure	Re-prompting when Claude says evidence is limited or conflicting	Ask Claude to explicitly flag uncertainty: "Note where evidence is weak, contested, or expert-opinion only." Uncertainty acknowledgment is a feature, not a bug.

The iteration principle: Context engineering is not a one-shot skill. Start with your best context. Read the response critically. Add what was missing. Correct what was wrong. The second prompt in a conversation is almost always better than the first — because now you have evidence of what Claude does and doesn't know about your situation.

Ready-to-Use ObGyn Contexts

Copy any context template below. Fill the bracketed fields with your specific clinical situation. Each one is engineered to give Claude the full context it needs to reason at the level of an experienced clinician — not a generic one.

Prompt engineering rule for clinicians: The more clinical context you front-load (gestational age, comorbidities, prior results, clinical setting), the more precise the output. AI quality is input quality.

Self-Assessment: AI Workflow Literacy

10 questions. Tests your ability to match task to workflow, use Claude features effectively, and avoid common mistakes. No data leaves your browser.

Quick Reference

Workflow Decision Guide

Your task	Best mode	Key tip
Quick DDx or drug question	Focused Session	Include GA, comorbidities, recent labs
Reviewing an uploaded guideline PDF	Workspace	Upload the PDF, then ask specific Qs
Manuscript editing or co-authoring	Projects	Include citation format + style examples
Journal peer review	Projects	Store your criteria + past reviews
Resident teaching case	Workspace or Projects	Upload curriculum objectives
Patient handout draft	Focused Session	Specify reading level (7th–8th grade)
Literature synthesis / grant background	Deep Research	Enable web search; allow 20–30 min
Statistical analysis of a dataset	Workspace	Upload CSV/Excel; ask for specific tests
Comparing two protocols	Workspace	Upload both docs in same session
Building a clinical calculator	Focused Session + Artifacts	Ask for single-file HTML output

Claude Features at a Glance

Feature	What it does	When to use	Access
Extended Thinking	Deep step-by-step reasoning	Complex diagnosis, guideline critique	Pro
Web Search	Real-time internet access	Recent trials, drug alerts, news	Free + Pro
Document Upload	Read PDFs, Word, Excel, images	Guideline review, data analysis	Free + Pro
Artifacts	Generates docs/code/tools in preview	Patient handouts, calculators, slides	Free + Pro
Projects	Persistent instructions + files	Recurring workflows, research programs	Pro
Memory	Remembers you across chats	Style preferences, specialty context	Pro
Deep Research	Multi-source research report	Systematic reviews, grant backgrounds	Pro
Skills (in Projects)	Saved behavior templates	Recurring task types	Pro

The 5 Context Engineering Rules

Front-load clinical context: GA, comorbidities, setting, prior results
Specify the output format you need: table, list, paragraph, template
State your audience: clinician, patient (reading level), resident
Ask for uncertainty to be flagged: "note where evidence is weak or contested"
For citations: "Vancouver format; flag any reference you cannot verify"

The one rule that matters most: AI quality is not model quality — it is context quality. A well-prompted Claude Sonnet outperforms a poorly prompted Opus every time.