Let me preface this blog post with a very clear and direct message. Do not do what I did. This is a bad use of generative AI. This is pure silliness with no real practical value whatsoever. This is a really, really, bad idea. But it was fun as hell, so here goes.

Last year I did two investigations into recipe parsing on the web. As we all know, most recipe sites go out of their way to make the actual recipe, you know, the thing you want to read, obfuscated and buried beneath a lot of stuff that is... well not the actual recipe. I first investigated JSON-LD and using that to parse web recipes into data: Scraping Recipes Using Node.js, Pipedream, and JSON-LD. This worked really well. I then followed it up with a more generic approach using Gemini and generative AI: Using Generative AI to Parse Web Pages into Data. This worked pretty good as well, and could be adapted for other purposes and types of content.

This came up in conversation with my friends today, and on a whim, because, this is how my 50+ year brain works, I wondered if it was possible to take a nice, filtered out recipe and go the other way? I.e., fill it with crap and obfuscate the simple recipe into a SEO-ridden nightmare and wall of text?

Yes, the answer is yes, you can. Let's get started, shall we?

I began by prototyping in Gemini's AI Studio. I used this as a system instruction:

You take a simple recipe and do your best to expand upon it with 
superfluous text, including a made up story about the author's 
childhood and why the recipe means so much to them. You should 
make the actual recipe incredibly difficult to find.

I then took this recipe my buddy Scott shared with me:

And the result was spectacular:

Beautiful. I could have stopped there, but then my buddy Todd had a suggestion - I should also add a bunch of snarky comments from folks asking for modifications, complainig, and so forth. And for fun, a few random political ones as well.

That sounded great. So I modified my system instruction to include:

Also return ten to twenty comments from slight angry, snarky commentors 
demanding different versions, like a vegan one, or one that will work on
Mars. the comments should not be helpful in anyway. Within the comments, 
sprinkle a few that are political and completely out of context.

The results were delicious:

That last one was perfect.

The final step I took was to define a structured output such that I'd get a JSON result with the recipe and an array of comments with name and the comment text. In case you're curious, this is how that looks in AI Studio:

I then took the code output from AI Studio and dropped into a Cloudflare Worker. All this does is look for a recipe string sent in as input, pass it to Gemini, and return the result. I'm breaking my "all Python" rule today as I haven't yet tried building a Python Cloudflare Worker, but want to do so soon.

const {
  GoogleGenerativeAI,
  HarmCategory,
  HarmBlockThreshold,
} = require("@google/generative-ai");

export default {
	async fetch(request, env, ctx) {

		const genAI = new GoogleGenerativeAI(env.GEMINI_API_KEY);

		const schema = {
			type: "object",
			properties: {
			recipe: {
				type: "string"
			},
			comments: {
				type: "array",
				items: {
				type: "object",
				properties: {
					name: {
					type: "string"
					},
					comment: {
					type: "string"
					}
				},
				required: [
					"name",
					"comment"
				]
				}
			}
			},
			required: [
			"recipe",
			"comments"
			]
		};

		const model = genAI.getGenerativeModel({
			model: "gemini-2.0-flash-exp",
			systemInstruction: "You take a simple recipe and do your best to expand upon it with superfluous text, including a made up story about the author's childhood and why the recipe means so much to them. You should make the actual recipe incredibly difficult to find. Also return ten to twenty comments from slight angry, snarky commentors demanding different versions, like a vegan one, or one that will work on Mars. the comments should not be helpful in anyway. Within the comments, sprinkle a few that are political and completely out of context.\n",
			generationConfig: {
				responseMimeType: "application/json",
				responseSchema: schema
			}
		});

		const corsHeaders = {
			"Access-Control-Allow-Origin": "*",
			"Access-Control-Allow-Methods": "GET,HEAD,POST,OPTIONS",
			"Access-Control-Max-Age": "86400",
	    };

		let { recipe } = await request.json();

		let result = JSON.parse((await model.generateContent(recipe)).response.text());

		return Response.json(result, {
			headers: {
				'Content-Type':'application/json;charset=UTF-8',
				'Access-Control-Allow-Origin':'*',
				'Access-Control-Allow-Methods':'GET'
			}
		});

	},
};

I then deployed, and because I wanted to see it live, I built a demo in CodePen using Alpine.js. Now, this is using my key on the Cloudflare side so I'll have to shut it down if it gets too much traffic, but for now, have some fun with it.

See the Pen Evil SEO Recipe Machine by Raymond Camden (@cfjedimaster) on CodePen.

Enjoy!