Cooking Without Entering the Kitchen

I am from an Afrikaans family and grew up in South Africa. Afrikaans culture is deeply agricultural in heritage. While I never lived on a farm, my mother spent a good part of her life on one, and the food-centric lifestyle associated with the "plaas" (farm) shaped large parts of what I value and who I am.

In my family there is a strong tradition of keeping and sharing hand-written recipes. This was an every-day form of connection for the generations of my great-grandmother, grandmother, and even still now my mother. Before everyone goes home after gathering for Christmas, the tannies and oumas (aunties and grannies) would exchange books and copy by hand the latest new recipes. Every girl had a personal recipe book by the time she left the house.

Through a series of inheritances, my mother now has the recipe books across three generations on both my parents side, stretching all the way up to my maternal great grandmothers and down to recipes we developed together when I was 8 years old. This treasure of culture, food and familial connection is priceless. But sadly, also borderline incomprehensible to me.

Cursive archaic Afrikaans, sometimes written in short-hand, on stained aging paper is sadly not very user friendly or accessible. When reading one of the recipes with my mother, I remarked that the only way I'd recognise a particular squiggly scribble as "Worcestershire sauce" is perhaps by inferring it from tasting the dark brown blotch obscuring half of the squiggle. To make matters worse, these books are all back in South Africa, and I am in London.

I wanted these recipes to be accessible, not just to my cursive-illiterate eye, but also to non-Afrikaans speakers and the wider world. Today Afrikaans cuisine isn't very accessible, even in South Africa where few restaurants do truly local cuisine. So I decided I would see if Claude could read them, translate them, and give me a nice, digitally preserved, legible copy in both English and Afrikaans.

I discussed the idea one evening on a phone call with my mother and she sent me a phone scan of a ginger-beer recipe. Three weeks later I had a polished web app capable of ingesting recipes and displaying them in a user friendly interface organised by book and author.

A few ways exist to get recipes into the app. My mother can capture images in the browser with her phone, upload existing photos, or connect it to her OneDrive and bulk upload PDFs of entire recipe books that she scanned.

Once an image arrives, a small assembly line of AI models process the recipes. One model reads the original Afrikaans squiggles, shorthand, stains and all. Another translates it into English. The final one then disassembles the result into a structured JSON file: ingredients list, method steps, quantity values, units of measurement, and what category it fits into. Each recipe then waits in a queue for a human (my mother) to glance over it, edit if needed, click approve, and maybe publish it for the world to cook.

What you end up with is each recipe on a standard page, with a tap to switch language, swap cups for grams, or scale the whole thing up for a crowd. There's a cook mode with large type that keeps the screen awake while your hands are covered in flour. And, the original photograph is always accessible with a tap, preserving my great-grandmother's handwriting. The recipes are organised by book and author, so you can wander through by whose hand wrote them. Uploading recipes is invite-only for the family at the moment, but soon I will be launching publically. More on that at the end.

I'm really not a coder. I know my way around querying a SQL database, and I took a Python course in uni, but let's be real, I'm not here to become a software engineer. I'm here to bake a pudding.

When I asked Claude to help me gather some facts for this piece, it informed me that across the entire build I had never once written a single line of code myself. It's right. I only read the code for the most part. So, what did I do? What role did I play in building this monument of culinary code?

Well, I'm also not a total layman. I work as a strategy consultant primarily in the AI space, helping large companies design AI visions, nailing down product roadmaps, and filling various product manager roles across different stages in the AI/tech use-case lifecycle. I know what the process and workflow of a highly skilled software engineering team looks like. I've spent plenty of time translating senior leadership's thoughts, feelings, frustrations and ideas into a concrete set of requirements for the dev team to go build. So this time it was just my own thoughts and ideas I needed to translate into requirements for Claude.

I primarily acted as the product manager and used Claude as a dev team that I could iterate the planning with before giving approval. Eyeballing it, I'd say my time and energy was split 70/10/20 between planning and aligning key decisions (70%), actually executing the plan to build the app (10%), and bug fixes (20%).

A simplified version of the process I followed:

All of the above documentation was given a home and maintained by Claude as we went. As I put it at the time, this would "become the Bible for the build effort going forward." Other than a mid-point restructuring exercise, I didn't have to touch the documentation very much at all, as I had given Claude instructions to build documentation maintenance into the protocol for each sprint's closeout. I.e. Claude updated all the relevant documentation after each change or decision made during the build.

I immediately broke my own rules and threw the sprint sequence out the window. Sort of.

I decided to build a proof of concept first, to test my AI pipeline design and ensure it worked the way I intended. A secondary goal, was to iterate with different models to keep API costs as low as possible. If it ended up costing $5 per recipe, this wouldn't be very workable.

The reasoning was that the pipeline was the only part I was unsure about. Everything else, the website, the database, the login, the file storage, is well-stewed beef. So I sliced just the pipeline sprint out of the roadmap cake and had Claude build it on its own, with no web server, no database and no login. Files went in from a folder on my laptop, and the results came out into a basic viewer I could open in a browser. The point was to answer one question before building anything around it: can the models read my family's handwriting well enough to be useful?

The core idea held up. With a decent prompt, the pipeline was reading recipes correctly on the first try, including some I struggled to read myself. That result let me cut two things I'd initially planned:

1. An "integrity check" step to catch nonsensical output, but it mostly flagged things that weren't actually wrong. E.g. an old recipe not stating an oven temperature isn't an AI pipeline error, it's just how my ouma wrote ("everyone knows you bake a cake at 180°C").
2. A fallback to a more powerful model for when the first pass failed, but the first pass rarely failed, so I shelved it.

The second thing I did was optimise the per recipe cost. I asked Claude to run a small experiment: take 16 recipes, and run every step of the pipeline through three different models, a cheap one (Haiku), a mid one (Sonnet), and the expensive one (Opus), saving the outputs side by side so I could compare them. Each step forked off of an upstream output from Opus so that errors wouldn't compound.

The whole experiment cost about $8 to run. Before giving my own opinion, I asked Claude to review the outputs and tell me what it thought, because I didn't want to lead it toward the answer I'd already half-formed. Its read and mine mostly agreed, though I overruled it on one step, it preferred Sonnet for parsing (which was the most token-heavy step) and I thought Haiku was fine.

The outcome was a tiered cake. Uuuh, I mean, tiered pipeline: Opus only for reading the handwriting, Sonnet to translate, and Haiku to structure the result. That brought the cost down to about $0.16 per recipe and around 26 seconds start to finish.

One exchange from this stage I found funny. Claude proudly flagged a case where Opus had confidently read an ingredient as "16 tsp baking powder" and pointed to it as an overconfident misread. I agreed, but later I checked with my mother. It was 16 teaspoons, the model was right, and the "sensible" reading would have been wrong.

Once I'd confirmed that what I was trying to achieve was feasible I decided I could move on to actually building the whole thing.

With the pipeline proven, the rest was the bread-and-butter stock standard app build, but standard still takes time, and anyone who's tried their hand at sourdough or churning butter can tell you that neither is quick and easy. This is where most of the 60 hours went.

I worked through it sprint by sprint, and a rhythm emerged that I more or less repeated each time:

If I build another app this way (and I plan to) I'll be following the above approach again. Roughly summarised, this turns into:

Building this has been a joy. I learned so much but also moved incredibly fast. It's hard to overstate the satisfaction of finding 20 bugs, giving the exact details and my hypothesis on why they might exist to Claude, and then 15 minutes later they're all fixed and my hypotheses are either proven or disproven.

I spent a good hour onboarding my parents onto the tool and during that we identified a dozen or so small features/bugs that needed work. I got Claude to address those and an hour later it was updated.

My main reflection from this isn't that coding has now become "easy" or a skill that's no longer "needed." Being a software developer is still definitely a needed and critical skill in the grand scheme of things. Rather, it's a scale and accessibility shift. With a small enough scale and simple enough ambition, it's now possible to be a product manager without a dev team. Which is a sentence that probably gives a lot of developers and engineers nightmares. Not from a "I'm losing my job to AI" perspective, but from a "Oh God, I hope I never have to touch that vibe-coded-spaghetti."

It's also been a truly great opportunity to connect with my parents too. We've had so many video calls laughing at the ridiculous things the earlier versions of the AI pipeline came up with (Haiku is truly terrible at Afrikaans, but unintentionally hilarious btw).

Building, testing and using this app has resulted in more conversation and quality time with my parents than I have had with them in months for any other reason. Hearing the stories of some of these recipes has been incredibly special.

One story, that I found particularly funny, is the story of a pudding that was labelled "Arina Poeding (duur een)" translating to "Arina Pudding (expensive one)". In the 80s, back in the small farming community where my mother grew up, there was an older tannie (auntie) in the community who did a lot of catering and bake-sales. She had developed her own low-cost version of 'Malva Pudding', a famous South African dessert. It was a huge hit and she shared this recipe fairly broadly, including to my Grandmother and a friend of her's, Arina.

One day, Arina found that she had lost the recipe. She reached out to the older tannie, and asked if she could copy it from her again. By this point, the older lady felt that since the recipe was so popular, she was owed her fair share. And so, she told Arina that if she wanted the recipe it would cost her thirty Rand ($90 today, if converting to USD and then carrying forward for inflation). An expensive pudding! Rather than pay the extortionate amount, Arina simply copied the recipe from my grandmother's recipe book instead. And ever since then, it has been known in my mother's family as the expensive pudding.

At one point, my grandmother added some dates to the pudding. When my grandfather tried it, impressed by the upgrade, he remarked, "Now you can sell this recipe for fifty Rand!" ($150 today)

Soon, I'll be launching the public version of the app, which I've decided to name Malva Recipes, to give others the chance to also preserve and share their handwritten family recipes in the same way. If you're interested, you can go sign up for the waitlist now at malva.recipes.