What started as building a personal taste.md skill for myself, turned into building a pipeline to create any taste as a skill. The most important piece is references. This is where you should spend time. If the references suck, so does the skill. I find that references cropped tightly... on details in high resolution work the best. Each image gets analyzed by both Opus 4.7 and GPT 5.5. The analysis is based on why the reference is successful as a piece of design - not what it does functionally. Using two models helps rule out biases and gaps from each. The models focus on layout, spacing, typography, rhythm, composition, hierarchy, etc. At the end, each image has: reference-01/ - opus-4-7-analysis.md - gpt-5-5-analysis.md Then we fuse them together using GPT 5.5 - but the md files are anonymized so 5.5 doesn't prefer itself. reference-01/ - fused-analysis.md reference-02/ - fused-analysis.md etc. After fusion, we have one synthesized analysis per reference. Now the goal is to combine all of those into a single rule set. This is where chunking matters. If you ask one model to combine 100 image analyses at once, the result becomes too broad. It summarizes instead of preserving the granular design rules we want. Instead we chunk the fused analyses into smaller groups. Each group gets merged into a chunk-level synthesis, usually from around 6 to 8 image notes at a time. Then one final model pass fuses those chunks into a single md rule set. Finally, using the rule set, we write a skill of concrete instructions. It enforces constraints, uses imperative wording, and avoids vague taste words.show more

Connect your OpenRouter account and try it here or clone the repo and run it locally with your own keys

I do something similar, the difference is I manually annotate the things I like on each reference before feeding to the skill.

Yeah that’d probably help here too

Dude this is super interesting, thanks for sharing the breakdown. I’m actually planning to write next weeks’s newsletter issue on the need to start collecting more references outside of text and this fits in nicely. 2 questions for you - 1. do you feel like it actually deducts correctly *why* you like something or what makes it good? You mention the “why the reference is successful” but it doesn’t seem like you actually tell it yourself why it’s significant to you? 2. do you feel like this has the potential to broadly capture your taste or is it more about finding a “local valley” in the model’s style that’s aligned with the particular references you provided (and so it’s more like style presets you can tune and switch between vs a larger overarching taste)

When you ask it to identify why it’s successful and ignore the function of the reference- it is surprisingly close to what I would say about each reference myself. I think you need to do this for each general aesthetic you’re looking for. Not broadly my complete taste as a whole

Makes sense. Thanks for elaborating

this is awesome! thanks for sharing. I tested with one sample and tested on both opus 4.7 and gpt 5.5 and results are amazing. I created reference sets so i can work on targeted taste experiments. 1 - local ui wrapper 2 - reference and results

Another banger

🫶🏼

that's lowkey genius af, thanks for making this 🫡

Super cool! I’m exploring something similar with Claude skills right now! trying to build a system/infrastructure that help anyone transform their taste / aesthetic into something can be understood, executed by AI for future projects and interfaces.

🔥 this is great for static refs, but one of the roadblocks I’ve hit in similar experiments is that motion is a big part of why I like what I like and none of the leading vision models can see video. hi-res, HFR sequences are usually too big for the context window. any tips?

there’s a few skills that help. @emilkowalski has a good one but overall still have to spend time finessing motion in general with the models

Better yet, use the Mobbin MCP server and give it thousands of references. Cool idea!

this is very cool

Curious what differences Opus and GPT notice? Is there much duplication from their analysis ?

They both find different things GPT seems to do better from a traditional design perspective

this is so cool!

Super great @Jaytel - love this space. We have done a similar thing for design systems that keeps enterprise styles consistent across. Will try this one 👊

super interesting stuff

Tried something similar (nowhere near that scale), pushing the model to use semiotics as its lens. Adds real depth beyond pure form.

Loved it! Starting with those aesthetics can save so much time and tokens!

Super cool, will give this a go today.

gonna try this 🫡

Brilliant!

This is so smart, great explanation too thank you 🙌

Love it, thank you so much for sharing. I will try to connect it directly to the Arena API instead and see if I can make it work that way as well, could be quite powerful!

Very interesting workflow, and if you are using Pi gotta try

Incredible

thank you for your work this is amazing

this is dope

Super intresting, I was thinking on this problem, loved the same image multiple times model call thing

Really nice design! Well done

Any more effective than “make it sexy?”

skill + “make it sexy” ?

Interesting, I am doing this the other way around, where I am feeding the model all the principles about different styles and building a skill from it so it can refer to it.

Nicee!

Multi-model analysis + chunking = surprisingly scalable design system.

This is super inspiring! Do you envision the skill evolves as your personal tastes change? How would the files capture that?

This is cool. Have you tried fine tune a model for this?

Hey, @Jaytel been playing w. this a bit. Only one model fired in my trial but still got a result. I made some tweaks that maybe are stearing things in the wrong way. It feels more like forced application of images refs rather than an interpretation of an overall aesthetic 🤔.