Prompt Engineering for Sensitive Medical History Documentation

Many LGBTQ+ people need to document medical history, gender dysphoria, or transition care in formal contexts: updating insurance, preparing for surgery, establishing care with new providers, or supporting legal proceedings. These documents require specific clinical language, appropriate medical framing, and precise detail—but writing them yourself can feel retraumatizing, and many AI outputs feel either dismissively clinical or inappropriately casual.

The core challenge is that AI tools respond to vagueness with generic templates that don't capture your actual situation. A medical history document for top surgery requires different emphasis than documentation for hormone therapy access, which differs from documentation for name change legal proceedings. The precision you need comes from structured prompting: techniques that guide the AI toward your specific context.

Structured prompt architecture

Effective medical documentation prompts follow a pattern: (1) role specification, (2) context constraint, (3) output format prescription, (4) tone/register instruction, (5) specific detail requirement. Compare weak versus strong approaches:

Weak: "Write a medical history."

Strong: "You are helping me draft a medical documentation letter for [specific clinical context]. My situation is: [your specific facts]. The letter must be suitable for [specific audience: insurance company / surgical provider / therapist]. Include these specific elements: [list them]. Use clinical terminology appropriate for medical professionals. Avoid mentioning [specific details you want excluded]. The tone should be [professional/empathetic/matter-of-fact]."

Example: Hormone therapy documentation

A trans person working with their doctor to access hormone therapy might prompt Claude: "I need documentation of gender dysphoria for my endocrinologist's records. I'm seeking testosterone therapy. My experience with gender dysphoria includes: [specific examples of persistent distress relating to sex characteristics/social role]. I've been consistently identifying as [identity] since [timeline]. My social transition status is: [details]. Format this as a clinical summary suitable for an endocrinologist's medical record. Use DSM-5 gender dysphoria diagnostic criteria as a framework. Include the diagnostic criteria you're addressing and how my situation aligns with each."

This structure produces documentation that feels authentically clinically grounded rather than self-described. The AI anchors its output to diagnostic criteria, which both increases credibility and ensures comprehensive coverage.

Privacy and data handling

Sensitive medical information should not be entered into ChatGPT's free tier, which may train on your inputs. Use Claude (which has stronger privacy commitments), Claude within restricted enterprise environments, or local AI models if handling detailed personal medical history. For preliminary drafting with less sensitive details, ChatGPT works fine. For final versions containing specific diagnoses, medical history, or identifiable information, Claude is preferable.

Consider decomposing sensitive information: draft a medical summary without your actual name or specific timeline details, then add personalization in a separate secure document.

Iteration for calibration

Medical documentation rarely lands perfectly on the first draft. Use iterative refinement: after receiving the AI's initial output, identify what feels misaligned (too clinical, missing crucial context, emphasizing wrong elements), then prompt for revision: "Revise this to better reflect [specific aspect]. Emphasize [what matters more]. De-emphasize [what matters less]. Change the tone to be more [specific adjustment]."

Try this: Draft a medical history summary for one specific purpose (like hormone therapy access). Write a detailed, structured prompt including your specific clinical context, the intended audience, and 3-4 specific elements you want included. Generate the output in Claude. Then revise it once based on what felt off about the first version. Notice how specificity in your prompt translates to relevance in the output.