How Kickstart Picks Your 50 Recipes (It's Not a Grab Bag)

Kickstart looks simple from the outside: answer some questions about how you cook, get 50 recipes as plain-text files you own. A questionnaire and a download button.

I want to show what happens between those two steps, because it's the part of Cook I've probably rewritten the most — and because "we'll pick recipes for you" is one of those promises that's easy to make and easy to do badly.

The naive version is terrible

The obvious implementation is a filter: take everything tagged with your cuisines, drop what hits your allergies, return the top 50. I know it's terrible because that's roughly where the first version started.

You get packs like this: eleven steak dinners, because steak recipes are popular and you said you like American food. Four carbonaras — technically different recipes! Breakfast is all smoothies, because a smoothie takes three minutes and you asked for quick mornings. There's a smoker recipe in there, and you don't own a smoker.

Every one of those recipes matches your answers. The pack is useless. That gap — between fifty individually-correct picks and one genuinely good collection — is where all the actual work lives.

It reads the recipes before it ever meets you

Every recipe in the pool (about sixteen thousand, all real and sourced, none generated) has been read and annotated ahead of time — not just cuisine and cook time, but the things you'd notice only by cooking it: whether it survives the freezer, whether it's genuinely one-pot or just claims to be, whether a kid will eat it, how much protein it really carries, which ingredients would trip a low-FODMAP diet, and what dish it fundamentally is — so the system knows a ribeye and a T-bone are the same idea wearing different names.

That last one matters more than it sounds. Tags can make two recipes look different when any cook can see they're the same dinner.

Your answers aren't treated equally

The questionnaire's answers get sorted into hard lines and preferences, and the matcher treats them very differently.

Allergies are absolute — no shellfish means no shellfish, full stop. But most answers aren't like that. If you don't own a grill, grill recipes don't vanish; they slide down the ranking, because a grilled chicken thigh survives a cast-iron pan just fine — while a missing pressure cooker really is disqualifying. Saying you dislike blue cheese doesn't cost you every cheese dish. And "I want to save time" caps how ambitious the pack gets, but not so far that breakfast collapses into drinks — a real quick breakfast is eggs in twenty-five minutes, not juice in three.

Getting those distinctions wrong is how recommenders end up feeling literal-minded. Most of the tuning has gone into exactly this.

The 50 slots are laid out like a week of eating before any recipe is chosen: so many breakfasts, so many dinners, a small side of desserts if you asked for them. Dinner slots hold dinners — sauces, sides, and drinks compete for their own little section instead of elbowing out a main course.

Then, as the pack fills, each pick changes what the matcher looks for next. Choose a rich braise and it starts leaning away from more rich braises — toward a different protein, a different method, a different weeknight mood. There are caps that stop any one cuisine or any one dish from colonising the pack, however well it scores. The goal isn't fifty recipes that each resemble your answers; it's a collection with the variety of a good cookbook.

It bends before it breaks

Some combinations of answers are genuinely tight — a strict diet, several allergies, and twenty-minute dinners doesn't leave sixteen thousand options. When the pool gets thin, the matcher relaxes its own diversity rules first — the gentlest constraints, never the hard lines — because a slightly uneven pack of fifty beats a perfectly balanced pack of thirty-one. Every request is logged with what was relaxed and why, and I read the odd-looking ones. That feedback loop is most of how it got smart.

Why not just ask an AI to write 50 recipes?

Because you'd get recipes nobody has ever cooked. Generated recipes read plausibly and fail in the pan — quantities that don't add up, steps that skip the part where dinner happens. Kickstart's intelligence goes into selection: understanding real recipes deeply enough to choose the right fifty for you. The things it picks have been cooked, by real people, many times.

(There is AI in Cook — CookBot plans your week from the collection you own — but it plans from real recipes too. Here's what that looks like.)

Try it

Kickstart is free — answer the questions, get your 50 as plain-text .cook files that are yours to keep, edit, and sync anywhere. If the pack doesn't feel like your kitchen, I'd genuinely like to hear about it: that's the signal the matcher learns from.