My Month-Long Recipe Workshop
Even Recipes Require a Human in the Loop
I spent a month developing a single-serving persimmon dessert for my sister’s birthday. Not because I’m slow, but because I know my creative process needs incubation time and my first ideas are usually wrong. I worked with three AI models—Claude, GPT, and Gemini Flash—discussing everything from sugar temperatures to Korean culinary traditions to whether butterflies should be viewed from above or in profile.
This isn’t a story about AI replacing human expertise. It’s about why “give me a persimmon dessert recipe” may be counterproductive, while sustained collaboration between a user’s embodied knowledge and AI analytical capability produces something genuinely unique.
AI as a Recipe Refining Sounding Board
I didn’t ask AI for a recipe from scratch. I found a YouTube recipe by Joconde as my jumping-off point, then brought it to my thinking A.I.des for analysis and identified the following problems.
The recipe used massive amounts of gelatin across multiple layers, creating a monotonous wobbly texture. It sieved the persimmon pulp, destroying the sacs—the best part texturally. It paired persimmons with cream cheese, which I feared might overpower their delicate flavor. The cookie base (store-bought Digestive biscuits) had nothing to do with persimmons or Korean culinary tradition.
That conversation made me articulate what I actually loved about hongsi (super-ripe persimmons)—those sacs with their inimitable texture, the challenge of making them digestively friendly, the desire to concentrate the best parts without the usual binding consequences of eating whole ripe persimmons.
Once I knew what mattered, AI helped me think through solutions:
Flavor pairings: Why walnuts counteract persimmon’s binding effects (traditional Korean wisdom), why yogurt’s tang would balance sweetness better than cream cheese, why sujeonggwa spicing (cinnamon and ginger) would evoke Korean tradition without descending into pumpkin-spice territory. GPT made the sujeonggwa (cinnamon-flavored punch) connection, which became the conceptual anchor.
Technical troubleshooting: When I couldn’t decide between walnut shortbread, speculoos, or gangjeong (candied walnut brittle) for the base, we worked through butter ratios, structural requirements, and flavor competition. When my gangjeong base dissolved into puddle, my thinking A.I.des explained sugar’s hygroscopic nature, which causes it to actively pull moisture from anything nearby.
Presentation concepts: GPT suggested visual ideas I’d never have imagined, including that hanji (Korean paper) scroll concept using translucent persimmon streaks to evoke traditional Korean paper’s fibrous texture. Genuinely sophisticated and culturally grounded, even though I ultimately went simpler for a personal gift.
Why the Human in the Loop Is Vital
None of them—not one, across dozens of turns—caught the thirsty sugar problem. I was building a dessert with candied walnut brittle base topped with moist persimmon sacs. Because of the sugar in the base, the brittle dissolved into puddle.
If I were a “prompt and pray” user who’d asked for a complete recipe and followed it blindly on the day of, I’d have discovered this failure when presenting the dessert to the recipient. Instead, I caught it during testing because I actually made the different versions and reported the results back to my thinking A.I.des.
The Trash Can Lesson
What made me finally understand sugar’s thirst? A random memory evoked by Claude’s explanation on sugar’s hygroscopicity: months ago, I’d thrown away candies I didn’t want to eat into my food waste bin. Days later, they’d turned into puddle from the moisture in my veggie peelings. At the time, I just thought “huh, weird.” But when Claude named the mechanism—“sugar doesn’t just passively absorb humidity, it actively pulls moisture through osmotic pressure”—I connected the two experiences.
Even my garbage teaches me food science. And it took AI explaining the chemistry for me to recognize that discarded candy as a relevant data point.
The Double-Candying Discovery
Through iterative testing, I discovered something genuinely novel: double-candying nuts creates moisture-resistant coating. The first candy layer seals the nut, the second provides structural binding. Even when outer syrup dissolves a bit from contact with moisture, inner coating protects the nut’s crunch.
This emerged from systematic experimentation—trying different barrier layers (persimmon leather, almond butter), testing various configurations, documenting what worked and why. AI helped me understand mechanisms, but the discovery came from actually making multiple versions and paying attention to results.
The Methodology I’d Recommend
Start with an existing recipe as a reaction point: Don’t ask AI for a recipe from scratch. Find something close, then analyze together what needs improving.
Iterate through conversation: My base went through multiple versions—walnut shortbread (too buttery), speculoos (wrong texture), finally candied walnut brittle with sujeonggwa spicing. Each version taught me something.
Test in real life: I discovered sugar temperature matters, that 118°C/244°F works for my needs, that double-candying creates moisture-resistant coating, that persimmon leather blocks moisture but makes slicing impossible while almond butter allows some syrup loss but preserves crunch. You only learn this by making it.
Use multiple models strategically: GPT excels at (verbally describing) visual brainstorming and creative suggestions, Gemini explains scientific mechanisms, and Claude provides analytical refinement and synthesis and is the “foodie” of the three. Triangulating across all three surfaces better solutions than any single model.
Recognize AI limitations: They’re pattern-matching across text, not tasting food or understanding ingredients through embodied experience. They can’t tell you (before you think to ask) moisture will dissolve sugar over time, or that oligosaccharide-coated sesame seeds look like ants, or that fruit leather ruins clean slicing. Only actual testing reveals these gaps.
What This Produced
A dessert with hand-separated persimmon sacs suspended in yogurt cream (unsweetened extra-thick Greek yogurt enriched with heavy cream), double-candied walnut brittle base with sujeonggwa-inspired spicing, almond butter moisture barrier, and simple sac-ring decoration. It solves the digestive concerns around ripe persimmons through yogurt probiotics and walnut pairing. It preserves the best textural elements rather than sieving them away. It’s grounded in Korean tradition while being completely original.
The recipient didn’t react dramatically—we have different tastes. But I got a month of fascinating problem-solving, learned about sugar chemistry and moisture barriers, and developed a recipe that could work in fine dining settings with those hanji scroll presentation GPT suggested.
The process was the reward.
The Larger Point
This is what engaged AI use looks like: human expertise directing collaboration, AI providing analytical support and creative brainstorming, real-world testing validating or rejecting ideas, iteration producing refinement that neither human nor AI could achieve alone.
“Prompt and pray”—asking AI for complete recipes and following blindly—is like asking a very well-read person who’s never cooked to write you a recipe. They’ll produce something that sounds plausible based on patterns they’ve seen, but it won’t account for moisture migration, highlight what makes ingredients special and how to best showcase them in a recipe, or solve problems they can’t taste.
Sometimes the most valuable thing AI can do is help you articulate what you already intuited—like realizing those persimmon sacs are the crown jewel worth building an entire dessert around. And sometimes the wisest thing you can do is ignore AI entirely and trust your gut when coated sesame seeds look like ants.
AI is a thinking partner for people who already know how to think. It’s not a replacement for expertise, embodied knowledge, or the willingness to actually make the food and see what happens.
[This post was drafted with assistance from Claude Sonnet 4.5, and informed by discussions with ChatGPT-5 and Claude Sonnet 4.5, and Gemini 2.5 Flash.]
GPT-5
Systematic Analysis of Recipe and Key Ingredients
Prompt: I’m brainstorming a birthday cake recipe. I have about a month to decide on a recipe and practice until the big day. The recipient LOVES persimmons, especially the super ripe ones that Koreans call hongsi. I found one recipe on YouTube when I search for “persimmon cake.” It’s by Joconde. Here are the ingredients:
♥︎Cookie Base
70g Digestive
1g cinnamon powder
25g melted butter
♥︎Walnut, persimmon layer
50g dried persimmon
18g walnut
♥︎Persimmon Puree
270g frozen ripe persimmon
53g sugar
10g lemon juice
30g orange juice
5g gelatin
♥︎Cheese Cream Filling
165g cream cheese
90g plain yogurt
100g 50% whipped cream
8g lemon juice
5g gelatin
♥︎Cheese Cream Layer
250g cheese cream filling
1. Because the birthday person is not a big eater, I’m going to make a miniature cake (single serving) and am going to use fewer layers.
2. They’re not a big cream cheese fan, so I’m just going to go with a yogurt cream. In a previous macaron experiment, I tested various brands of extra thick Greek yogurt. Used it as a macaron filling to balance out the very sugary macaron shells that don’t leave much room for creative adjustments. That same yogurt could be used in my yogurt cream. I will have to see if it needs the gelatin to set up or if I can just use it as is. I don’t want to mix in sugar into the yogurt because it creates a slightly more fluid consistency and a translucent look that I found offputting.
3. I’m also not a fan of using store-bought cookies in the base, and am wondering why Joconde did not use a crushed candied walnut base for the cake instead. Walnuts and persimmons are classic pairings because these super-ripe persimmons are like cheese (not conducive to “regularity”), while walnuts counteract that effect.
4. The recipe seems to use a LOT of gelatin. I guess it’s ok because you do have a textural contrast in the base (which is going to be walnuts in my case), but it could probably use a bit more contrast and much less gelatin.
4.1. How about bits of candied dried persimmon on top instead of the persimmon jelly dome with whipped cream in the Joconde recipe? Thinking of candied ginger bites that you get in fruit cakes, etc.
4.2. The walnut bottom could be made like the top of a walnut/pecan pie (so it’d be like an upside-down walnut “cake”).
4.3. My favorite part of super-ripe persimmons is the “sac,” which has an even better texture than any jelly, so I’ll be separating these and using them in the “cake” and puree only the rest of the flesh for where I need the persimmon flavor. The sieved puree is probably the other big flaw in Joconde’s recipe because it doesn’t maximize all the attributes of a persimmon vs. some other sweet fruit.
5. Wondering if nutmeg might be a better fit for persimmon. Not many people know that nutmeg works really well with dairy (even mac and cheese), but I do.
Claude Sonnet 4.5
Traditional Candied Nut Base
Prompt: Yes, and hongsi is local and seasonal to boot. Following our discussion, I collected some recipes for gingersnaps, gingerbread cookies, and speculoos (because the other two required molasses, which I couldn’t get right away). The speculoos came out ok but a little caky, probably because I made them a little thick and also because the recipe said 325F, when it should have been 350F for the crispy speculoos in the accompanying picture. Flavor-wise, it works with the persimmon and yogurt cream. Not as distracting as that walnut shortbread crust. But then today I realized I’d missed a painfully obvious alternative: the Korean candied nuts/sesame seeds (강정). These can be shaped, don’t contain distracting butter, and stay crispy!
Spice Infusion for the Syrup
Prompt: I thought of a further refinement for my persimmon dessert (this is my process; iteration until I get something I feel really good about). Instead of mixing in the ground spices into the water for the syrup, how about infusing the water with the spices (and getting rid of the particles) and using that brew for the syrup instead?
A Recipe with Finesse
Prompt: I think a fine dining establishment might like this recipe since they have people willing to stage just for the experience and the name on that resume. Diners would appreciate it, too, especially those who were too afraid to eat hongsi because of the “regularity problem.” And you get the best part in high concentration on a plate just for you! That’s a dessert that makes you feel like you’re being pampered in the best possible way (might evoke those memories as a child when your mom carefully deboned your fish for you, as in a Korean version of the movie Ratatouille).
That Thirsty Villain Sugar!
Prompt: Still interesting that none of the AI I discussed the recipe with caught the thirsty sugar problem. If I have to cast ingredients as cartoon characters, sugar would be a thirsty villain.
Doubling Down May Be the Way to Go
Prompt: I tried both the leather and the almond (I figured it might stand out less than pistachio here) spread on the backup dessert I’d saved. I layered the leather under one half of the dessert and the almond spread under the other half, since I had only one backup. Didn’t try much of a sloping effect, just a 1mm coat of the spread and two thin pieces of the leather. The structure of the leather must have been dense enough to keep any moisture from the sacs and yogurt from dissolving the base, which wasn’t yielding to the fork when I tried to cut through it (past the cream layer and the leather). But once I set those aside, I could easily cut that base. That half had stayed whole but still cuttable. And the walnut pieces were crunchy!
The other half under the almond butter layer had oozed quite a bit of syrup. But the nuts were still crunchy!
What I think happened here: they stayed crispy because they were double-candied. These were nuts that I had overtoasted, coated with syrup, but set aside for experiments because I feel guilty about throwing out food (candy isn’t food, though, so I’m ok binning it if I don’t like it). Then yesterday, I cut those candied walnuts and coated them with fresh syrup to get discs of even and reasonable thickness for my base. So like those double-fried Krispy Kreme donuts, double-candying might be the way to go in my future experiments where I have to use a brittle/gangjeong base (with an almond butter buffer, of course)?
