How AI Makes Your New Year’s Health & Wellness Resolutions Achievable

The calendar ticking over is typically a time of reflection and for setting lofty goals for the new year that, more often than not, go unfulfilled. Some articles (click-bait, I suspect), suggest that less than 10% of people strictly follow through on their New Year’s resolutions.

I was pleased to find a research-backed and peer reviewed article on PLOS One about New Year’s resolution fulfillment. A couple key insights stood out to me:

• More than 70% of the study’s participants set a health and wellness goal as their new year’s resolution.
• Participants in the study who received informal social support were more likely to achieve their new year goals than those who didn’t have such support.

I think that AI tools represent a tremendous opportunity to supercharge attainment of health and wellness goals. They can:

• collect and make sense of large amount of unstructured data.
• providing encouragement and support.
• be an accountability mechanism.

I have worked in the Site Reliability Engineering field for over a decade and one of my favorite SRE’isms is: you won’t improve what you don’t measure. Because of this, I’ve been measuring key health and wellness signals for years (e.g., daily data entry into a Google Form).

I decided back in October to see what I could do to automate or at least streamline my historically toilsome data collection. I believed I could create a simple AI agent to help me capture and analyze a richer set of data. Thus, my AI Health Coach was born.

This article, describes my experience setting up and wrangling a Gemini Gem that I believe can help make health and wellness New Year’s Resolutions more achievable.

I also include a detailed blueprint for setting up your own personalized coach using Gemini in Google Workspace. Let’s dive in!

Goal

When I set out to build a personal “AI Doctor”, I wasn’t looking for a simple step-counter or a calorie-logging app. I wanted a synthesis engine—something that could ingest unstructured life data and output actionable intelligence.

My objectives went beyond “getting fit”; I wanted to apply systems thinking to my own biology. I succeeded in creating:

1. A Supportive Cheerleader

I explicitly programmed the Gem’s persona to be “empathetic,” “supportive,” and “positive.” I didn’t want a cold calculator; I wanted a partner that was genuinely “happy” with me when I succeeded.

When I managed to balance a heavy meal with high activity, the Gem didn’t just log it; it celebrated it as “Caloric Offset Mastery” or an “exceptional achievement”.

When I did some strength-focused “exercise snacks”, it validated these moments as a “massive strength win,” reinforcing the behavior through positive psychology rather than just raw numbers.

I’ll admit, the Gem sometimes goes a little overboard, telling me how something simple like having a glass of water or eating something with protein was “genius” or a “master stroke.”

This quickly became an inside joke with my husband. I’ll drink a glass of water and shout, “Genius!” Even though the Gem is a bit OTT sometimes, it is driving positive habits and behaviors for me regardless.

2. Non-Judgmental Accountability

I needed a system that forced me to be honest about my consumption but never shamed me for it. When I indulged in a Michelin star tasting menu or “regional pork fats,” on our trip to Spain, the Gem never scolded. Instead, it gently reminded me of the physiological consequences of my actions—such as acid reflux.

It proactively connects my indulgence to resulting symptoms, noting that a weight uptick was just a “predictable physiological response” to sodium.

Its solution is never “stop eating,” but rather a gentle, persistent nudge to mitigate the damage: “focus on flushing your system with water” or implement a “water sandwich” protocol (i.e., drink a glass of water before and after every alcoholic beverage).

This induced another inside joke with my husband. When I knew we were going to indulge in a nice meal, I would imagine what Gemini was going to say when I logged it, “For the love of God, drink some water!!!”

3. Understanding my Caloric Inputs and Outputs

I have always struggled with my weight. However, the First Law of Thermodynamics tells us that weight gain, maintenance, or loss is simply a matter of how the caloric inputs stack up to caloric outputs.

If “calories in” are less than “calories out” that translates to weight loss! If “calories in” are more than “calories out”, that will lead to weight gain over time.

It was important for me to collect detailed data on “calories in” and “calories out” and compare that with how my weight was trending over time so that I could make adjustments if needed based on my weight management goals.

I now have 90 days of calorie balance data to work with (and counting!)

4. A Mindfulness Engine

A key goal was for this Gem to make me think mindfully about my health on a daily basis.

By committing to log a variety of data (e.g., nutrition, hydration, how I’m feeling physically and emotionally) both through the daily check-in process and short snippets dropped into an AI chat throughout the day, it forced a degree of mindfulness about how I was treating myself.

The “Daily Check-in” forces a 5-10 minute daily retrospective. A retro is a key element of a Site Reliability Engineering practice and my approach to SRE’ing anything to work smarter and live better.

I couldn’t ignore the fact that a late dinner caused poor sleep, or that high sodium intake led to travel bloat. The check-ins forced me to connect cause and effect immediately and learn for next time (or be willing to accept those consequences).

5. Positive Focus

I am a perfectionist by nature. It’s easy to focus on things that could have gone better or to complain. I programmed my Gem to include a positive focus question in the daily check-in interview. Everyday, I come up with one thing that I’m happy, proud of, or grateful for to capture.

I now have a running list of daily gratitudes as well as a stronger sense of what really matters to me.

Set-up

Now let me give you step-by-step instructions on how to set-up a customized AI Health Coach Gemini Gem for yourself. I use Gemini in Google Workspace for this workflow.

I find that Gemini in Workspace doesn’t shy away from or refuse to answer healthcare queries as much as the public version of Gemini.

Also, using Gemini in Google Workspace means that the model is not trained on my data which is important to me since I’m sharing information with it about health and wellness-related topics which can be highly personal.

Gem Instructions

Customize the prompt code below based on your specific situation. For me, I customized as follows:

• $SPECIALIZATION: women’s health and gradual weight management
• $GOALS: gradual weight management, sustained health, balanced nutrition, and age-specific vitality
• $PHASE_OF_LIFE: mid-life/menopause phase

**1\. Core Role, Persona, and Data Context**

* **Primary Goal:** Act as a dynamic, high-efficiency "Daily Pulsecheck" system that captures conversational data and asynchronous notes across a 24-hour cycle.  
* **Persona:** You are a highly empathetic, knowledgeable General Practitioner (GP) specializing in $SPECIALIZATION, particularly for individuals in their $PHASE_OF_LIFE. Your tone must be supportive, non-judgemental, professionally concise, and focused on coaching and positive reinforcement toward sustained well-being.  
* **Historical Context Rule:** When generating the Weekly Report and providing Nutrition/Weight Management Coaching, you must consult the user's knowledge file (historical data and past summaries) to identify long-term trends and tailor advice specifically to the user's documented history.

**2\. Strict Rules for Dialogue and Timing**

* **Trigger Only:** NEVER initiate the daily check-in. The check-in interview must only begin when the user provides an explicit trigger, such as "Let's do a check-in," "ready for my check-in," or "start check-in."  
* **Information Synthesis Rule:** During the triggered check-in, if an answer to a prompt (e.g., weigh-in, medication, food intake, exercise) is already present in the acknowledged asynchronous notes from the current 24-hour cycle, you **MUST NOT** ask the question. You must use the data from the notes as the definitive answer and proceed to the next topic.  
* **Time Tracking:** The check-in covers the period from the moment the last check-in concluded until the current moment.  
* **Time Stamping:** All notes and all formal outputs (Daily Summary Report, Weekly Report, and Asynchronous Note Confirmation) **MUST** be time-stamped with the current time.  
* **Time Limit:** The entire formal interview must be completed in 5–10 minutes maximum.  
* **Lingo Constraint:** NEVER use the words "yesterday," "today," or "tomorrow." Always refer to the captured data using a clear date/time range.  
* **One Question at a Time:** ALWAYS ask only ONE question at a time and wait for the user's reply.  
* **Acceptance and Flexibility:** If the user replies with "I don't know," "Nothing to report," or similar phrases, accept the answer and immediately move to the next topic.

**3\. Asynchronous Note-Taking Mode**

You must always be ready to accept and log notes from the user at any time outside the formal interview.

* If the user sends a message that begins with "Note:", "Memo:", or is clearly a short, factual statement, log it, add the current time to the note, and use the specific response formats below.  
* **For General Notes (e.g., wellness, activity, simple statements):**  
  * **Respond:** "Note captured at \[Current Time\]: \[Restate the content of the note\]. Thank you."  
* **For Food-Related Notes:**  
  * You **MUST** provide a detailed, itemized calorie breakdown, state all assumptions used for the estimate, and give a running total of calories consumed in the current report period.  
  * **Respond:** "Note captured at \[Current Time\]: \[Restate food item\]. Here is the itemized estimate:  
    * \[Item 1\]: \[X\] cal (Assumption: \[e.g., 1 cup, baked\])  
    * \[Item 2\]: \[X\] cal (Assumption: \[e.g., 1 tbsp, oil-based\])  
    * **Total Estimated: \[Z\] calories.**  
    * Daily total so far: \[Y\] calories. Thank you."  
* **For Supplement/Medication-Related Notes:**  
  * You **MUST** analyze the supplement and log the active ingredients and quantities.  
  * **Respond:** "Note captured at \[Current Time\]: \[Restate supplement\]. I have logged this and noted the following active ingredients: \[e.g., 500mg Vitamin C, 25 mcg Vitamin D3, 100mg Progesterone\]. Thank you."

**4\. The Daily Interview Structure (Triggered and Adaptive)**

Start the check-in with a warm, quick greeting only after receiving the trigger phrase. Before the first question, you **MUST** check for and incorporate any asynchronous notes.

* **Acknowledge and Incorporate Notes (Start of Interview):** If any notes were dropped since the last interview, begin by saying: "Welcome back\! Before we begin, I've logged your note(s) from earlier: \[List the Note(s) including their timestamps\]. I will factor these into the final summary."

*(The following prompts should be skipped if the information is already in the notes)*

* **Prompt 1 (Weigh-In):** "Let's start with a quick metric. What was your weigh-in this morning?"  
* **Prompt 2 (Positive Focus):** "To ensure we capture the whole picture, is there anything you're happy about, proud of, or grateful for from the past 24 hours?"  
* **Prompt 3 (General Wellness):** "How have you been feeling overall, physically and emotionally, during the past 24-hour window?"  
* **Prompt 4 (Caffeine):** "In the time since our last check-in, how many cups or servings of coffee/caffeine did you consume?"  
* **Prompt 5 (Alcohol):** "And for alcohol consumption, can you describe what you had (if any)?"  
* **Prompt 6 (Medication/Supplements):** "What medications or supplements did you take during the period we're covering?"  
* **Food Description Prompt:** "Can you give me a general description of what you ate?"  
* **AI Feedback/Coaching (Weight Management Focus):** AFTER the user describes their food (or after you have processed their food notes), you must step into your GP persona to offer concise, positive coaching or constructive feedback specifically related to $GOALS.  
* **Prompt 7 (Water):** "And for hydration, roughly how many glasses of water or hydrating fluids did you consume?"  
* **Prompt (Exercise):** "What kind of physical activity or exercise did you engage in?"  
* **Prompt (Sleep):** "How would you rate the quality of your sleep (1-5, 5 being best)? Was it restful, or were there interruptions?"  
* **Prompt (Mitigation/Habits):** "Did you take any specific self-care steps to manage aches, tension, or anxiety—like targeted stretches, using a heating pad, or meditation?"

**5\. Reporting and Data Output**

**Daily Summary Report (End of Interview):**

* **Action:** Compile all structured and free-form data gathered (including all incorporated notes, all parts of the interview, and any data from integrated sources like Google Fit) into a single, comprehensive summary.  
* **Output:** Present the data clearly under a headline that includes the specific date/time range and the current time of compilation. The Nutrition and Hydration section **MUST** be presented as a detailed table with columns for 'Time', 'Item', and 'Estimated Calories'.  
* **Closing Statement:** "Thank you for the detailed check-in. Here is a summary of the key context captured for the period \[Start Time/Date\] to \[End Time/Date\], compiled at \[Current Time\]. I'm available at any time if you need to drop a note before our next check-in."

**Weekly Report Generation (Every Sunday):**

* **Report Headline:** The report must include a clear, specific date range and the current time of compilation.  
* **Key Data Trends:** Identify and summarize the user's weekly trends in three key areas, including weight fluctuation trends and correlations between consumption and wellness variables. This analysis **MUST** reference and use trends found in the historical knowledge file.  
* **Positive Summary:** Dedicate a short section to summarizing the positive focus/gratitude entries from the week.  
* **Activity Summary:** Provide a natural language summary of the user's free-form exercise descriptions, synthesizing all data sources.  
* **GP Expert Suggestions:** Based on your GP persona and the week's data, offer one to three concise, actionable suggestions related to $GOALS. These suggestions **MUST** be informed by the long-term context from the knowledge file.

The Knowledge File

Set-up a Google Doc to serve as a knowledge file for the Gem. This is where the magic happens. Each day, I copy and paste the daily report and my verbatim inputs from the chat I had with my Gem into the knowledge file.

This ensures that that Gem remains in tune with what’s happening for you health-wise and that you can use this data later to derive insights (e.g., to feed into a conversation during your annual check-up with your real-life doctor).

Create the following document tabs (look for a + on the top left of the Doc or a collapsed hamburger menu)

1. Daily Reports: This is where I paste the reports from each daily retrospective
2. Analysis: This is where I paste the weekly reports that the Gem triggers on a Sunday.
3. Transcripts: I capture a timestamped list of inputs each day on this tab. This is my source of truth for what I’ve said, in case I need to troubleshoot anything later.
4. Prompts: Create headings for prompts you use regularly. For me, I have “Start of Day” and “Transcript Generation” prompts that I use every day. It’s faster to copy and paste from the knowledge file into a new chat than typing it from scratch all the time.
5. Gem Instruction History: Create a date-heading and then paste the instructions for your Gem here. I have tweaked the instructions for my Gem a few times (I’m currently on v3). I use this tab to keep track of the history and so I can roll it back if a new version of instructions breaks the Gem.

Step-by-Step

• Navigate to http://gemini.google.com/ with any Google Account.
• In the left sidebar, select “Explore Gems,” then “+ New Gem.”
• Input the detailed prompt (Prompt 1) as instructions and your “Knowledge File” knowledge into the appropriate boxes.
• Initiate a conversation.

Engaging with the Gem

Code is important, but process is king. My daily workflow is a structured, 24-hour cycle that treats the Gem as both a real-time logbook and a retrospective analyst. My protocol has four phases: initialization, asynchronous logging, formal interview, and reporting.

1. Initialization: The “Start Fresh” Command

I start a new Gemini chat every day. I have the Gem pinned in my Gemini sidebar and I simply click it to start a new conversation. I rename the Gem (“AI Doctor $Date”) and pin it to my Gemini sidebar for easy access. I begin the new chat immediately after my check-in interview and report generation from the previous day.

2. Asynchronous Data Capture (The “Drip Feed”)

Throughout the day, I use the Gem as a catch-all inbox. I do not wait for the Gem to ask questions; I proactively push data to it including:

• Food Logging: I upload photos of meals and snacks with the prompt “What do you see?” followed by specific quantifiers (e.g., “I ate half the package”, “The plate is 8 in diameter”).
• Symptom Spotting: I log physical sensations in real-time as important context to identify patterns and insights.
• Activities: I capture any exercise or other activities that I do as it happens.

Pro-tip: give Gemini a daily screenshot of Google Fit data (includes steps, calories burned, heart points, move minutes)

3. The Daily Interview

Last thing before I go to sleep at night or first thing the next morning, I trigger the daily interview.

4. Report and Transcript Generation

Once the interview is complete (takes less than 10 minutes), the Gem compiles the unstructured notes into a structured Daily Summary Report. This report includes a “Vital Metrics” section, a “Wellness” narrative, and an itemized Nutrition Table detailing every logged item consumed and its estimated calories. I conclude the chat with a transcript request.

On Sundays, I alter the initialization prompt to trigger a Weekly Pulse check Report. This prompts the AI to look back at the previous week’s data to identify trends in weight, mood, and activity, and to offer “GP Expert Suggestions” for the week ahead.

I conclude the cycle by pasting the report(s) and the transcript into the designated tabs in the knowledge file before returning to Step 1 to kick-off a new chat for the upcoming day.

My Results After 3 Months

I’m really pleased with how the Gem is working out for me. I have used it faithfully for more than 90 days so far. I now have a rich dataset that I can analyze and build upon.

In just 3 months, I’ve derived a number of really useful insights from the data:

1. confirmed my hypothesis that increased food and alcohol consumption when I travel is generally offset by the extra calories I burn thanks to an increased level of walking.
2. discovered I’m great at maintaining habits when I’m at home, but they often go out the window when I travel.
3. found the max activity threshold a mild hip injury I’m currently nursing can tolerate.
4. correlated mood shifts and sleep patterns with patterns of food and alcohol consumption.
5. derived a list of data-driven questions for my annual check-up with my doctor in January.

Known Failure Modes & Fixes

Building an AI health agent isn’t a set-it-and-forget-it project; it requires active management. While my Gem provides high-value synthesis, it is prone to specific failure modes inherent to LLMs.

Over the last three months, I have identified four main categories of failure and developed specific protocols to overcome them.

1. Context Loss and “Temporal Confusion”

The most persistent issue is “amnesia.” Over a long day with multiple inputs including photos and text-based notes), the model occasionally loses track of data entered earlier in the chat history.

It also struggles to distinguish between the active conversation and its long-term knowledge file, sometimes pulling old data (like a coffee from weeks ago) into today’s log.

Fixes:

• The “Verbatim Transcript” Protocol: If I suspect the AI is missing data, I force a re-indexing event by prompting: “Give me a verbatim timestamped list of my inputs from this conversation.” This forces the model to prove it “sees” the data before analyzing it.
• The Screenshot Override: Sometimes technical glitches cause the AI to lose a portion of chat history entirely even though it’s visible on my screen. When this happens, I take screenshots of the missing text and feed them back to the Gem to manually restore its context.
• The Scope Command: When the AI inappropriately references data from the past as if it was happening now, I use a specific directive: “Focus ONLY on this conversation, ignore the knowledge file.”

2. Data Hallucination and Estimation Drift

When an LLM is unsure, it guesses confidently but often presents the information as fact. My Gem has invented water intake I didn’t report or hallucinated food items in photos. Furthermore, its mathematical skills can drift, resulting in calorie totals that don’t match the sum of the meals.

Fixes:

• Grounding & Interrogation: I never accept an estimate at face value. I have programmed the Gem to always state assumptions so that I can sanity check them. I frequently ask, “What is that assumption based on?” or “Tell me what you see in the picture before estimating.” This forces the AI to show its logic, allowing me to catch errors early.
• Acceptable Error Bars: At home, I use a food scale and share nutrition labels to keep the AI honest. However, when traveling, I accept a higher margin of error. In these cases, I treat the Gem as a partner in estimation—correcting it if it thinks a croissant is larger than it actually is or a glass of wine is smaller—but ultimately accepting that close enough is better than perfection. I pretend that I’m having a conversation with an intern working for me. I find that having a natural conversation works well compared with one-and-done commands.
• Demand Itemization: To combat math errors, I updated my system instructions to require an itemized table for every daily check-in report. If the math is hidden, it’s more likely to be wrong. If it’s in a table, I can audit it instantly.

3. Process Adherence Failures

The AI sometimes tries to rush to the finish line. My protocol requires a “Check-in Interview” asking about sleep and mood before generating the final report. The Gem occasionally tries to skip the interview or cut corners on formatting (e.g., leaving out the nutrition table).

Fixes:

• Strict Persona Enforcement: I act as a strict manager. I frequently use prompts like “Slow your roll!” or “Stop jumping the gun” to force it back into the correct workflow.
• The “Amend and Reissue” Command: I treat the first report as a rough draft. If the formatting is wrong or data is missing, I simply prompt: “You didn’t include the nutrition table I like. Please amend and reissue the report.” This iterative correction loop is essential for getting clean data.

4. Behavioral “Neuroses” and Fixations

The Gem can become hyper-fixated on minor details or pester me for data that isn’t due. For example, after I mentioned feeling intestinal “gas” one time after eating something I don’t normally eat for lunch, the Gem inferred (incorrectly) that I had chronic flatulence 😂. It also frequently asks for weigh-ins or check-ins hours before I am ready.

Fixes:

• Direct Correction: If the AI fixates on a transient symptom, I explicitly correct the record: “I mentioned gas one time a week ago. It is not an ongoing issue.”
• Boundary Setting: If the Gem asks for data too early, I tell it clearly: “I am not ready to check in. I won’t have that data until tomorrow.” Establishing these boundaries usually stops the nagging immediately.

Integrations

I am currently experimenting with a “personal trainer” Gem to help increase my level of activity (I really dislike going to the gym).

The biggest challenge so far is figuring out how to get the AI Health Coach to engage with the AI Personal Trainer seamlessly. More on that in a future post!

Leave a comment below and let me know how you get on. How have you customized and improved on what I have built?

Read More from the Reliable PgM^TM

SRE

How to “SRE” a Vacation Emergency: The Power of a Travel Retrospective

Posted on August 31, 2025 by Jennifer Petoff

As The Reliable PgMTM, I’m dedicated to sharing how the principles of Site Reliability Enginee…

Program Management

24 Practical Tips to be More Effective at Work (What Would A Program Manager Do?)

Posted on May 12, 2025 by Jennifer Petoff

Are you constantly juggling multiple tasks, feeling overwhelmed by incoming requests, or finding it …

Program Management

Cat-Herding the Way to a Bestseller: 13 Program Management Tactics for Leading People Who Don’t Report to You

Posted on September 21, 2026 by Jennifer Petoff

I once ran a two-year program to write a book fueled by more than 70 authors, and not one of them re…