> ## Content Index
> Fetch the complete content index at: https://blog.warrenweb.net/llms.txt
> Use this file to discover other available public pages before exploring further.

# 💭 AI Diabetes Coach—AI Prompt Engineering
- URL: https://blog.warrenweb.net/ai-coach-part-3/
- Published: 2026-04-07T00:42:26.000Z
- Updated: 2026-04-28T12:00:15.000Z
- Description: How I went from asking 'Why is my glucose high?' to getting personalized daily coaching that improved my time in range from 82% to 98%. Six iterations of prompt engineering that transformed generic AI into an essential health tool.
- Author: Doug Warren
- Tags: PKM, AI, Diabetes, Coach, Claude, Prompt Engineering, Health-Tech

---

**🤖 *AI Diabetes Coach* series:**  
*— 🚀 Part 1:* [The Transformation](https://blog.warrenweb.net/ai-coach-part-1)  
*— 🏗️ Part 2:* [System Architecture](https://blog.warrenweb.net/ai-coach-part-2)   
*➔ 💭 Part 3:* [Prompt Engineering](https://blog.warrenweb.net/ai-coach-part-3)   
*— 🕸️ Part 4:* [Knowledge Graph](https://blog.warrenweb.net/ai-coach-part-4)   
*—* 🔧 Part 5: [Build Your Own](https://blog.warrenweb.net/ai-coach-part-5/)   
*—* 🍴Part 6: [Teaching Your AI Coach to Cook](https://blog.warrenweb.net/ai-coach-part-6)

---

## The Generic AI Problem

*Why basic ChatGPT doesn't work for health coaching*

With diabetes (or any health issue), asking an AI a generic question gets you a generic answer—interesting perhaps, but rarely helpful. As they say, "Garbage in, garbage out".

For example, if I ask ChatGPT:

> Why is my glucose high?"

I get something like this:

> High blood glucose can happen for several reasons, depending on your situation (whether you have diabetes, what you’ve eaten, stress levels, etc.). Common causes include:  
>  
> **\- Diabetes** \- Your body can't regulate sugar properly- **Food** \- Carbs raise blood sugar- **Stress and illness** \- Hormones affect glucose- **Medications** \- Some drugs raise blood sugar- **Lack of sleep** \- Disrupts insulin sensitivity- **Inactivity** \- Exercise helps lower glucose- **Dawn phenomenon** \- Morning hormone surge  
> **What you can do right now:**Check your glucose againDrink waterTake a short walk if safeReview what you ate and when  
> *If you want, tell me your glucose reading(s), when you measured (fasting, after eating, etc.), and any symptoms... and I can help you narrow down the likely cause.*

So what does this tell me? Not much.

It's like asking "Why is the sky blue?" without any context about where I am, what time it is, or what I'm actually seeing. The response lists every possible reason the sky *might* appear blue, but doesn't help me understand "my specific sky" right now.

**What I actually need** is to understand my timelines, patterns, and trends. Not generic possibilities—but specific insights about *my* glucose, *my* meals, *my* insulin, and *my* daily patterns. That's what lets me learn and adapt to improve my health.

Generic AI can't do that. But engineered prompts can.

## Six Prompt Iterations

*How I evolved from questions to coaching*

### Iteration 1: Basic Questions

*No context, generic answers*

**What I tried:**  
Just asking health questions directly:

- "What should my glucose be after breakfast?"
- "Is 156 mg/dL too high?"
- "Why did I spike to 180?"

**What I got:**  
Generic medical ranges and boilerplate advice:

- "Normal post-meal glucose is 140-180 mg/dL"
- "Consult your doctor about target ranges"
- Spikes can be caused by many factors..."

**What was missing:**  
Any knowledge of ME:

- My Type 1 diabetes
- My insulin doses
- My meal timing
- My target ranges (70-160, not 140-180!)
- My patterns over time

**What I learned:**  
AI needs MY context, not generic medical facts.

### Iteration 2: Added My Diabetes Type

*What changed, why it mattered*

**What I tried:**

- "I have Type 1 diabetes. Why is my glucose high?"
- As someone with T1D, is 156 too high after breakfast?

**What I got:**  
Type 1-specific information—better than generic:

- References to insulin dependence
- Mentions of carb counting
- Tighter target ranges suggested

**What was missing:**  
Still no knowledge of MY specific situation:

- My current insulin regimen
- My actual breakfast and timing
- My glucose trend (rising? stable? falling?)
- My exercise plans or stress levels

**What I learned:**  
Diagnosis is important context, but I needed current data—not just historical facts about my condition.

### Iteration 3: Added Current Data  

**What I tried:**  
Including real-time information:

- "I have Type 1 diabetes. My glucose is 156 at 10 AM. I ate breakfast at 7:30 with 3u insulin. Why am I high?"

**What I got:**  
Much more specific analysis:

- Considered insulin-to-carb ratio
- Looked at 2-hour post-meal timing
- Suggested possible carb underestimation
- Mentioned dawn phenomenon possibility

**What was missing:**  
Context beyond this single moment:

- What did I eat for breakfast?
- How much did I actually carb-count?
- Is this high normal for me after this meal?
- How does today compare to yesterday?

**What I learned:**  
One data point tells a story, but patterns tell the truth.

### Iteration 4: Added Daily Pattern  

**What I tried:**  
Providing my full day's data:

- All glucose readings
- Every meal with carb counts
- All insulin doses with timing
- Exercise or unusual events

**What I got:**  
Pattern recognition within the day:

- "Your breakfast spike suggests under-dosing by \~1u"
- "Lunch was stable—that ratio worked well"
- "Evening low might be from afternoon exercise"

**What was missing:**  
Historical comparison:

- Is today typical or unusual?
- Which patterns repeat across days?
- What changed from yesterday to today?

**What I learned:**  
Daily patterns reveal cause and effect, but I needed to compare across days to see what's sustainable vs. what's a one-time event.

### Iteration 5: Added Weekly Context  

**What I tried:**  
Pasting several days worth of data:

- Monday through Friday glucose logs
- All meals and insulin doses
- Exercise and stress notes

**What I got:**  
Multi-day pattern analysis:

- "You consistently spike after breakfast tacos"
- "Your Friday evening lows correlate with yard work"
- "Weekend patterns differ from weekdays"

**What was missing:**  
This was incredibly powerful... but exhausting:

- Manually copying/pasting data each time
- No memory between conversations
- Starting from scratch every session
- Unsustainable long-term

**What I learned:**  
I needed automation and persistent memory, not just more data.

### Iteration 6: The Breakthrough  

**What I tried:**  
Three game-changing elements:

1. **Claude Projects** for persistent memory across conversations
2. **Automated data sync** via Dexcom API (glucose) + Glooko (insulin)
3. **Structured daily reviews** generated from my own data

**What I got:**  
True personalized coaching:

- Daily analysis of yesterday's patterns
- Comparisons to my historical trends
- Specific, actionable recommendations
- Conversational dialogue that builds on prior discussions
- No manual data entry required

**What was missing:**  
Nothing. This was the breakthrough.

**What I learned:**  
The combination of automation, memory, and structure transformed AI from "occasionally helpful" to "daily essential coaching."

**Result**  
82% → 98% time in range over 14 days. The proof that engineered prompts work.

## What Makes It Work

*Specific, context, examples, conversational*

Four principles emerged from my iterations:

### 1\. Specificity over Generality

Don't ask: "What should I eat for breakfast?"  
Ask: "I have Type 1 diabetes, my fasting glucose is 110, I have 2 hours before a morning workout— what's a good breakfast that won't spike me during exercise?"

Generic questions get generic answers. Specific questions get actionable guidance.

### 2\. Context over Instructions

Don't say: "Analyze my glucose".  
Say: "Here's today's glucose data, meals, and insulin doses. I spiked to 210 after lunch yesterday with the same meal but didn't today— what was different?"

Context lets AI see patterns you might miss.

### 3\. Examples over Explanations

Don't explain: "I want daily summaries".  
Show: "Here's the format I want:

- Average glucose: X
- Time in range: Y%
- Meals that worked well: Z
- Patterns to watch: ..."

Examples are clearer than descriptions.

### 4\. Conversation over Commands

Don't command: "Give me a meal plan".  
Converse: "Looking at my patterns this week, which meals kept me most stable?

Can we build on those?" Dialog builds understanding over time.

## The Results

*Real data from real days*

After implementing Iteration 6 (automated data sync + Claude Projects + daily reviews), my diabetes control transformed:

**Before engineered prompts:**

- Time in Range (Tight 70-160): 82%
- Inconsistent patterns
- Reactive management (correcting problems)
- Frustration with unexplained spikes and crashes

**After engineered prompts:**

- Time in Range (Tight 70-160): 98%
- Clear pattern recognition
- Proactive management (preventing problems)
- Understanding of my specific triggers

**14-day streak results:**

- Average Tight TIR: 92%
- Standard TIR (70-180): 97%
- Learned from every challenging day
- Recovered quickly from weekend restaurant meals
- Identified exercise timing effects
- Understood my roller coaster patterns

The difference wasn't just numbers—it was confidence. Instead of guessing why my glucose did something, I could trace cause and effect. Instead of generic advice, I got specific insights about MY body's patterns.

Engineered prompts turned AI from a search engine into a coach.

## Getting Started

*A progressive approach for your own prompts*

You don't need to jump straight to Iteration 6\. Start simple and add complexity as you learn what helps:

**Week 1: Start with specificity**

- Replace "Help me with diabetes" with "I have Type 1 diabetes, my fasting glucose is 110, what should I consider for breakfast?"
- Add your diagnosis, current situation, and specific question

**Week 2: Add context**

- Include your current data: glucose reading, timing, recent meal, insulin dose
- Ask about relationships: "What might explain this spike?"

**Week 3: Show patterns**

- Share 2-3 days of data
- Ask for pattern recognition
- Notice what insights emerge from comparison

**Week 4: Experiment with structure**

- Try different data formats
- See what Claude (or ChatGPT or Gemini) understands best
- Refine based on what gives useful responses

**Eventually: Automate what works**

- If manually entering data gets tedious, consider automation
- Use Claude Projects (or ChatGPT memory) for persistence
- Build on what you've learned through experimentation

**Remember:**

- Start where you are (no judgment!)
- Add one improvement at a time (iterate!)
- Keep what works, discard what doesn't (learn!)
- Share what you discover (help others!)

The goal isn't perfection—it's progress. Each iteration taught me something. Yours will too.

## What's Next

- 🕸️ **Part 4:** [Knowledge Graph](https://blog.warrenweb.net/ai-coach-part-4) — Giving AI persistent, queryable memory so it stops forgetting everything between sessions.
- 🔧 **Part 5:** [Build Your Own](https://blog.warrenweb.net/ai-coach-part-5/) — How to start building your own AI health system, regardless of technical background. This puts it all together with AI-powered analysis and the PKM services that tie the graph to your daily workflow.
- 🍴**Part 6:** [Teaching Your AI Coach to Cook](https://blog.warrenweb.net/ai-coach-part-6) — How food choices connect directly to glucose outcomes — building a recipe intelligence system that learns from your actual responses.

*Start from the beginning: 🚀 Part 1:* [*The Transformation*](https://blog.warrenweb.net/ai-coach-part-1)

The PKM structure that organizes all this data is covered in a separate series: 📅 [Day Cluster](https://blog.warrenweb.net/day-cluster) (Part 1).