WhatIsARecipe

What is a Generative Recipe?

A generative AI recipe is a schema for vibe coding. It is how you start without guessing.

You are not handing an AI a vibe and hoping. You are handing it a framework where the basic design decisions are already made: what this thing is, how it should feel, what it must do, and what it must not invent. The AI builds from that. You get a working prototype fast. Then you iterate the recipe until the artifact matches what you wanted.

The recipe takes the guesswork out of how to start.

Who should write one

At best, a recipe is written by someone with design experience.

Not “can I code this,” but “do I know the experience I want.” Intent, constraints, flow, emotional arc. The person who can say this should feel like a toy you pick up once and grin is more useful here than the person who can name a physics engine.

If you cannot name the feel and the must-nots, you do not have a recipe yet. You have a prompt.

What it is

• A structured description of intent, constraints, and feel
• A starting frame so the model does not have to invent the product
• A thing you iterate — same as a design doc, not a frozen spec
• Portable across models. Different AIs read the same schema and still build

What it is not

• Source code
• A prompt salad
• A guarantee of a finished product on the first pass
• The project owning what your AI generates. You run the model. You own the output.

The loop

1. Read the recipe.
2. Give it to an AI. Tell it to build from the schema and ask before adding extras.
3. Run the prototype.
4. Note what is wrong in the experience, not in the file tree.
5. Tighten the recipe. Build again.

Stop treating generated files as the source of truth. The recipe is the source of truth. The code is a rendering.

What a good recipe decides up front

Enough that a capable model should not stall on “what are we making.”

• Purpose
• Feel / emotional arc
• Hard constraints (platform, size, dependencies, single file, etc.)
• Must-haves
• Must-nots
• How you will know it worked

Field-by-field shape: [SchemaConventions].

How to run one: [HowToUse].

Worked example: [SquigglyCatapult].

Rule of thumb

If two different AIs can read the same recipe and both reach a playable or usable prototype in one pass, the recipe is doing its job. If they wander, the design decisions are still in your head. Put them on the page.