The 4 Rules
These apply to every AI interaction. Memorize them, live by them.
1. AI drafts. You verify.
Your review process is non-negotiable.
2. Confidence ≠ Correctness.
The most dangerous outputs sound the most certain.
3. When in doubt, leave it out.
Never share PHI, confidential data, or trade secrets.
4. You’re the driver.
AI is the car. You decide where to go.
AI Vocabulary
The terms you’ll hear most often, explained in plain language. Don’t memorize — reference.
LLM (Large Language Model)
The “brain” behind AI tools like ChatGPT, Claude, and Gemini. It predicts what words should come next based on patterns learned from billions of documents. It doesn’t “think” the way you do — it recognizes patterns.
Example: When you type a question, the LLM predicts the most likely helpful response based on everything it learned during training.
Prompt
What you type or say to the AI — your instruction, question, or request. The quality of your prompt directly affects the quality of the response.
Example: “Summarize this report in 3 bullet points” is a prompt. So is “Help me write an email to my team about the project delay.”
Context Window
How much information the AI can “see” at one time. Think of it as a desk — the AI can only work with what’s currently on the desk. When the desk gets full, older things fall off.
Example: Claude’s desk holds about 500 pages. ChatGPT-4’s holds about 300 pages. Gemini’s holds about 2,500 pages.
Hallucination
When AI confidently generates something that isn’t true. It might cite a study that doesn’t exist, invent a statistic, or create plausible-sounding but completely fictional information.
Example: AI might tell you there’s a 2024 FDA guidance on a topic that doesn’t exist — and it will sound completely confident doing so.
Bias
AI reflects the patterns in its training data — including assumptions, stereotypes, and blind spots. Your review catches what the model can’t.
Example: AI may make default assumptions about “typical” patients or healthcare providers, or use language that inadvertently excludes certain groups.
Token
A chunk of text that AI processes — roughly 4 characters or ¾ of a word. Context windows and pricing are measured in tokens.
Example: The word “pharmaceutical” is about 4 tokens. A typical email is 200–400 tokens.
Multimodal
AI that can work with more than just text — including images, audio, video, and documents.
Example: Uploading a photo of a whiteboard and asking AI to turn it into a formatted list is using multimodal capabilities.
Agent / Agentic AI
AI that can take actions on its own, not just generate text. Instead of just writing an email, an agent could write, format, and send it.
Example: An AI agent might research a topic, compile findings, create a presentation, and email it to your team — all from one request.
🔗 Clarity by Leanna: “Anatomy of an Agent” — This concept explored further
RAG (Retrieval-Augmented Generation)
A technique where AI pulls in outside information before generating a response. Helps reduce hallucinations.
Example: When you upload a PDF and ask questions about it, the AI is using RAG — retrieving information from your document to augment its response.
Fine-tuning
Customizing a general AI model for specific tasks or industries.
Example: A pharma company might fine-tune a model on regulatory documents so it better understands FDA submission requirements.
The Context Window — How AI “Remembers”
Think of it as a desk. The AI can only work with what’s on the desk right now. When the desk gets full, older things fall off.
What This Means for You
- Long conversations may “forget” early context — start fresh if the AI seems confused
- Share relevant documents in the same conversation where you need them
- If AI loses track, restate key points — put them back on the desk
- Personalization settings stay because they’re always loaded at the start
Hallucination — When AI Gets It Wrong
AI doesn’t know the difference between true and false — it only knows what sounds plausible. The most dangerous hallucinations sound the most certain.
Verification Checklist for ANY AI Output
- Can I verify this claim with an independent source?
- Does this citation actually exist? (Search for it separately)
- Does this statistic appear in any credible publication?
- Would I stake my professional reputation on this being accurate?
- If this is wrong, what are the consequences?
Bias — What AI Inherits and How to Catch It
AI learned from the internet, books, and articles. That content contains assumptions, stereotypes, and blind spots — and AI absorbed all of it.
Questions to Catch Bias
- Does this content make assumptions about who the reader is?
- Would this language feel inclusive to all intended audiences?
- Are there perspectives missing that should be represented?
- Does this reinforce stereotypes, even subtly?
- Have I asked AI to explicitly consider diverse perspectives?
🔗 Experience Is Everything by Barbara: “The Little Red Ferrari That Thought It Could” — Psychological safety with AI explored further
The Safe Swap Guide
If you can’t share the data, share the structure. Use placeholders to get the structure, tone, and format right, then swap in real details on your secure machine.
| ✘ Don’t Use This | ✔ Use This Instead |
|---|
| Patient Jane Doe, 67, Stage III breast cancer | A female patient in her 60s with advanced disease |
| Asset OWR-77, our Phase II oncology drug | A Phase II oncology candidate |
| Dr. Smith at Memorial Hospital, our lead KOL | A key opinion leader at a major cancer center |
| Q3 revenue of $47M, partnership with Roche | Strong quarterly revenue, major pharma partnership |
Never Put Into AI
Patient health information (PHI) • Confidential clinical trial data • Proprietary strategies or trade secrets • Non-public financial information • Employee personal information • Passwords or credentials • Pre-launch materials
You’re the Driver
When you learned to drive, nobody expected you to rebuild the engine first. The same is true for AI. Your job is to direct AI toward your goals, navigate around obstacles, and arrive at useful outcomes.
✔ Drivers
- • Know where they want to go
- • Direct AI toward their goals
- • Iterate until they get useful results
- • Stay in control of the process
- • Own the outcome
🚫 Passengers
- • Wait to be told what’s possible
- • Accept whatever output they get
- • Don’t know where they’re going
- • Feel at the mercy of the technology
- • Blame AI when things go wrong
The Key Players
Personalize Your AI
Generic AI gives generic results. When AI knows your role, industry, and communication style, every response starts closer to what you actually need.
Where to Find Personalization Settings
Voice Mirror Technique
Don’t guess how to describe yourself — let AI analyze your actual communication style. Gather 3–5 examples of your best professional writing, then use this prompt:
VOICE MIRROR TECHNIQUE
I'm going to share several examples of my professional writing. Please analyze them and create a comprehensive profile of my communication style that I can use to personalize AI tools. For each dimension below, describe what you observe:
VOICE & TONE - How formal or casual? Direct or diplomatic? Warm or reserved?
STRUCTURE - Bullet points, numbered lists, or prose? Lead with main point or build up?
VOCABULARY - Technical jargon or plain language? Distinctive patterns?
AUDIENCE AWARENESS - How do I address different stakeholders?
After your analysis, write a 4-6 sentence summary I can copy directly into my AI profile settings. Start with "I am a..." or "I work as a..."
===== MY WRITING SAMPLES BELOW =====
[Paste your 3-5 writing samples here]
Memory Storage Prompt
MEMORY STORAGE PROMPT
Please remember the following information about me for all our future conversations:
PROFESSIONAL CONTEXT
[Paste your Voice Mirror summary here]
ROLE & RESPONSIBILITIES
- My title/role: [Your title]
- My industry: [Biopharma / Life Sciences / Healthcare / etc.]
- Key responsibilities: [What you do day-to-day]
- Who I communicate with: [Your key stakeholders]
MY PREFERENCES
- I prefer responses that are: [concise / detailed / balanced]
- For formatting, I prefer: [bullet points / prose / numbered lists]
- My tone should be: [Your 3 voice words from the Voice Mirror]
Please confirm you've saved this information.