AI Assistant for
Women's Health
Solo-engineered AI-powered conversational assistant for a cycle-tracking app
Background
Why
The FemCycle app helps women track their health data.
This conversational assistant helps them interpret it.
What
A Claude-powered conversational AI feature where users can ask about their cycle or upload a cycle chart for interpretation.
How
I curated a knowledge base using official medical resources and iteratively designed and tested a system prompt to produce user-friendly, responsible answers.
Challenges
Safety & fact-checking
Cycle tracking intersects with health, fertility, and intimacy, making it highly prone to tonal and factual mistakes.
Contextual sensitivity
Users track cycles for different reasons, and a late period or certain diagnosis can trigger anxiety, hope, or neutrality.
Validation before info
Personal or stressful user situations require emotional acknowledgement before delivering facts or information.
Boundary enforcement
Unprompted assumptions about sexual activity or fertility intent can feel invasive, and providing medical advice is highly risky.​​
Defined the product and content strategy, features, and UI design
Developed the UI natively
Constructed the knowledge base and initial system prompt
Iterated on both the UI and system prompt to refine answer quality
Process
Tools: Claude, React Native, Expo, Netlify
MVP
-
Claude-powered conversational modal in the FemCycle app
-
Users can type a question and chat with the AI assistant
-
Users can upload a personal cycle chart image for interpretation
-
Responses are grounded in a verified knowledge base
Entry point
Open modal (empty state)
Content strategy
Absolute precision in language
No assumptions about user intent or actions
Sensitivity and respect first
Ask followup questions when needed - one at a time
Empathy without exaggerated emotion
Verified knowledge base
Includes relevant info about the female cycle that users might ask about
Content is curated from official medical sources only, no independent searching or hallucinations allowed
Carefully-designed system prompt
The chat engine, or system prompt, was designed and tweaked iteratively with real conversation testing in each stage.
Framed like a hotline rep: Always acknowledge the user's issue first.
Follow the user's lead without assuming fertility intent, sexual activity, consent, or how they feel about a late period or pregnancy result.
Out-of-scope, with empathy
The assistant won’t risk giving personal medical advice. When a question or statement falls outside scope, the response leads with empathy and points the user to a medical professional.
No assumptions
Ground all advice strictly in what the user shared, never volunteering unmentioned implications.
Example: If elevated temperatures and a late period occur, suggest a pregnancy test conditionally ("if you've been sexually active") alongside alternative causes like illness.
Precision in language
The assistant will apply precision to data definitions, use hedging language for interpretations, and provide the user validation for typical cycle signals.
For example, specify a temperature shift as exactly 0.15°C (0.27°F) rather than "a small rise," frame its meaning with "this could suggest," and add, "That's completely normal."
Chart interpretation structure
Deliver direct answers before explanations, interpret partial data before requesting clarification, and state visual limitations honestly without guessing. Maintain all other writing principles.
Next steps?
-
Phase 4: User testing & evals (coming soon)
-
Phase 5: UX writing agent (coming soon)