The Nutritionist's Dilemma
She knew the calorie count of every regional dish in southern Taiwan. Milkfish congee: 280 calories per bowl, but only if the cook used the traditional ratio of fish to rice. The night market version with extra sesame oil? Closer to 370. She had spent twelve years building this knowledge — clinic consultations, academic papers, her own meticulous measurements of local ingredients.
Her clients kept asking the same question: "Is there an app for this?"
There was not. The calorie apps on the market knew about chicken breast and brown rice. They did not know about milkfish congee. They did not know that a serving of Tainan-style coffin bread contains roughly 520 calories, or that the calorie difference between a handmade and factory-made pork floss rice ball is about 40%. Her knowledge existed only in her head and in a binder of handwritten notes.
She looked into hiring a developer. The quotes came back: $3,000 to $8,000 for a basic interactive tool, plus ongoing maintenance costs. For a solo nutritionist running a clinic, that was not a line item in the budget. It was a fantasy.
This is the expert's trap. You have knowledge that is genuinely valuable — specific, tested, hard-won. But the distance between "I know this" and "other people can use this" is a canyon, and the bridge across it has always been software development. A skill you do not have. A service you cannot afford.
That bridge just got shorter.
What You Are Actually Building
An interactive single-page web tool built entirely from your domain knowledge. Not a chatbot. Not a document. A real tool that other people can open in a browser and use.
For the nutritionist, this might be a calorie calculator that knows every dish at a Tainan night market. For a wine sommelier, a pairing recommendation engine that asks three questions and suggests three bottles. For a language teacher, a grammar checker that understands the specific mistakes Mandarin speakers make when writing English. For a fitness coach, a workout generator that accounts for common home equipment limitations in small apartments.
The key insight: the quality of the tool depends on the quality of your domain knowledge, not your coding skill. The AI handles the code. You handle the truth.
Think of it this way. A coding AI is like a very fast, very literal architect. It can build any structure you describe. But it cannot invent the floor plan — it does not know how many rooms you need, which direction the windows should face, or that the kitchen needs to be near the dining room because that is how your family lives. That knowledge is yours. The architect just makes it stand up.
Prerequisites
- Claude Code installed and authenticated. If you have never used it, the First Hour with Claude Code tutorial gets you from zero to working in under 60 minutes.
- A Claude Pro subscription ($20/month) or an Anthropic API key.
- Your domain knowledge. This is the most important ingredient. More on this shortly.
No coding experience required. Seriously. If you can type a paragraph, you can do this.
Step 1: Write Down What You Know (30 Minutes)
This is the step most people skip, and it is the reason most AI-generated tools are mediocre. They type "make me a calorie calculator" and get a generic calorie calculator that knows nothing about their specialty.
The AI is like a very literal but very fast intern. If you tell that intern "make a calorie calculator," you will get one — a basic, generic one pulled from textbook knowledge. If you instead hand that intern a 500-word document describing exactly what makes your calorie calculations different, you will get something nobody else has.
Open any text editor. Write a document that describes your expertise. Not what you want the tool to look like — what you know. Structure it like this:
1. Who uses this knowledge? Your clients, your students, your audience. What do they need? What questions do they always ask?
2. What do you know that generic tools get wrong? This is the gold. The specific, non-obvious facts that come from years of experience. For the nutritionist: the calorie differences between regional variations. For the wine sommelier: why a "red wine with red meat" rule fails for Taiwanese beef noodle soup.
3. What are the categories? List the major groupings. Dish types. Wine regions. Grammar error categories. Exercise types. The AI needs structure to build a usable interface.
4. What are the rules? The logic behind your recommendations. "If the client is vegetarian and training for a marathon, increase carb ratio to 60%." "If the wine budget is under $30, never recommend Burgundy — suggest these Chilean alternatives instead." These rules are your expertise distilled into if-then statements.
Here is what the nutritionist's document might look like:
# Southern Taiwan Regional Dish Calorie Database
## Audience
Health-conscious locals who eat traditional food daily.
They want to know what they are eating, not switch to salads.
## What generic calorie apps get wrong
- They list "fish congee" as one item. In southern Taiwan, milkfish
congee varies by 90+ calories depending on preparation method.
- Night market portions are 1.3-1.5x restaurant portions.
- Many dishes have a "dry" and "soup" version with very different
calorie counts (e.g., dry noodles vs soup noodles: 15-25% difference).
## Categories
- Breakfast: congee, egg crepes (danbing), soy milk combos
- Noodles: dry, soup, thick, thin, with/without meat
- Rice dishes: bento, rice balls, fried rice, congee
- Night market: skewers, fried items, soups, desserts
- Drinks: tea shops, traditional drinks, smoothies
## Rules
- Default portion = restaurant standard, not night market
- Always show range (min-max) because preparation varies
- Flag dishes where "healthy-sounding" is misleading
(e.g., vegetable tempura: 400+ cal due to batter and oil)
- Group by meal occasion, not food type — that is how people decide
This document does not mention HTML. It does not mention JavaScript. It does not describe what the app should look like. It describes what you know. That is the raw material.
Spend 30 minutes on this. Be specific. The more detail you put in, the better the tool becomes.
Step 2: Turn Knowledge into a Tool (10 Minutes)
Open your terminal. Navigate to a folder where you want the project to live.
mkdir my-calorie-tool && cd my-calorie-tool
claude
Now paste your knowledge document into Claude Code, followed by your instruction. Here is the pattern:
Here is my domain expertise on southern Taiwan regional dish calories:
[paste your entire knowledge document here]
Based on this knowledge, create a single-page interactive web tool
(one HTML file with embedded CSS and JavaScript) that lets users:
1. Browse dishes by meal occasion (breakfast, noodles, rice, night market, drinks)
2. See calorie ranges for each dish with notes on variations
3. Build a daily meal plan and see total calories
4. Flag misleading "healthy" options
Make it mobile-friendly. Use clean, modern design. All data should come
from the knowledge I provided — do not use generic calorie databases.
Press enter. Watch.
Claude Code will read your expertise, understand the structure, and start writing code. You will see it creating an HTML file with embedded CSS and JavaScript. The data — every dish, every calorie count, every rule about portion sizes — comes from your document. Not from the internet. Not from a generic database. From you.
This typically takes 2-5 minutes. When it finishes, you will have a file — probably called index.html.
Step 3: Open It and Test (5 Minutes)
open index.html
On macOS, this opens the file in your default browser. On Linux, use xdg-open index.html. On Windows, use start index.html.
You now have a working interactive tool in your browser. Click through it. Check the data against what you know. This is where your expertise matters most — you are the only person who can verify whether the tool is correct.
You will probably find things to fix. Maybe the calorie count for coffin bread is wrong. Maybe the categories need rearranging. Maybe you want a search function. Good. Go back to Claude Code and say:
The coffin bread should be 520 calories, not 480. Also, add a search
bar at the top that filters dishes by name. And add a "common combos"
section that shows typical meal combinations with total calories —
like "danbing + soy milk + egg = 485 cal".
Claude Code will update the file. Refresh your browser. Check again. Repeat until it matches your knowledge.
This iterative process is exactly how software development works, except you are doing it in plain language instead of code. Each round takes 1-3 minutes.
Step 4: Make It Yours (5 Minutes)
The file is on your computer. You own it completely. There are no subscriptions, no platform dependencies, no API costs after the initial creation. It is a file. You can:
- Email it to clients. They open it in any browser. It works offline.
- Host it for free on GitHub Pages, Netlify, or Vercel with one command.
- Edit it later by opening Claude Code again and describing what you want to change.
- Add more data as your knowledge grows.
To put it online so anyone with the link can use it:
npx netlify-cli deploy --prod --dir .
Follow the prompts. In under two minutes, you will have a public URL for your tool. Share it on your clinic's website, your social media, your business card.
Why This Works: Knowledge is the Bottleneck
Most people think the hard part of making a tool is the coding. It is not. The hard part is knowing what the tool should do — what data it needs, what rules it follows, what edge cases exist.
An AI can write code all day. What it cannot do is spend twelve years measuring the calorie content of regional Taiwanese dishes. It cannot do the ten thousand wine tastings. It cannot teach three thousand students and learn exactly which grammar mistakes keep recurring.
That expertise is the bottleneck, and you already have it.
The prompt engineering here is not clever tricks or magic words. It is a detailed, honest explanation of what you know. Think of it this way: prompt engineering is explaining a task to a very literal but very fast intern. The better your explanation, the better the result. And nobody can explain your domain better than you.
Beyond the First Version
Your tool does not have to stay simple. Once you have the basic version working, you can ask Claude Code for upgrades:
- "Add a comparison mode" — let users compare two dishes or two wines side by side.
- "Make it remember user preferences" — using browser local storage, no server needed.
- "Add a print-friendly version" — so clients can take a physical copy.
- "Support English and Chinese" — Claude Code can add language switching in one pass.
- "Add images" — describe where to put them, drop image files in the folder, done.
Each upgrade is another conversation with Claude Code. Describe what you want in plain language. Get the result. Check it against your expertise. Ship it.
Three More Examples
This is not limited to nutrition. The pattern works for any domain where you have structured knowledge.
Wine sommelier — pairing engine: Write down your pairing rules. Not "red with meat" — your actual rules. Which specific wines work with Sichuan hot pot and why. What to recommend when someone says "I like sweet wine but I do not want to seem unsophisticated." The awkward real-world questions that no generic wine app handles.
Language teacher — error checker: Document the twenty most common grammar mistakes your students make. Not generic English errors — the specific ones that speakers of your students' native language produce. Include example sentences, corrections, and the explanation you give in class. The AI turns this into a tool where students paste their writing and get feedback in your voice.
Fitness coach — workout generator: List your exercise library, with the equipment each one requires. Document your programming rules: how you sequence exercises, how you adjust for beginners versus intermediate, what substitutions you make when someone does not have a pull-up bar. The tool asks three questions and produces a workout plan that sounds like you wrote it — because you did, in the form of rules.
What You Learned
- Your expertise is the product. The AI is the manufacturing process. Do not skip the knowledge document.
- Specificity beats cleverness. A 500-word description of what you know produces a better tool than a 50-word "clever" prompt.
- Iteration is normal. The first version will not be perfect. The third version will be good. The fifth will be something you are proud to share.
- You own everything. No subscriptions. No platform lock-in. One HTML file that works in any browser, forever.
The gap between "I know something valuable" and "other people can use it" used to require a developer, a budget, and months of work. Now it requires a text document and an afternoon.
Your knowledge deserves to be a tool. Go build it.