You write a prompt, you get an image. It's good. It's glossy. But it isn't yours. And it looks exactly like everyone else's.
The reason is simple. AI learned from millions of images and gives you their average. And the average has no taste.
Picture a new designer on their first day with you. Tell them "I want clean, but not boring" and they won't get it. Show them 20 pieces you love and 10 you can't stand, and within an hour they will. AI works the same way.
The system in 5 steps
- Show. Images you love, and images you really don't.
- Pattern. Claude finds what exactly pulls you in.
- Blind test. New images. Claude predicts whether you'll like them.
- Correct. Where it was wrong, you explain. Every mistake becomes a new rule.
- Skill. Your Taste Card becomes a skill, and every new visual request runs through it.
The secret isn't one prompt. The secret is the system you build once, which then works for you.
Prep: 20 minutes
Before you open Claude, gather your material. This is the most important part: bad references never produce good taste.
01What to collect
Claude learns your taste
Open a new chat in Claude. Fill in the brackets with your own details.
This prompt turns Claude into your taste coach. It tells you how many images to send, in what order, and when to correct it.
Role: You are my visual taste coach. You help me teach you my taste through examples, not descriptions.
Goal: Learn my visual taste well enough to predict what I like before I tell you.
Context: I create [Instagram posts / product photos / ads / carousels] for [who]. I make images with [Midjourney / ChatGPT images / other]. I am tired of glossy, generic AI images.
Process you lead:
1) Tell me exactly how many references to send and in what order: likes first, then dislikes, then "almost" images.
2) After each batch, ask me up to 3 short questions about what I notice.
3) Do not guess my taste from one image. Wait for patterns.
4) When you have enough, tell me, and write a Taste Card.
5) Run a blind test before we turn it into a skill.
Rules: be honest, say when you are unsure, never flatter my references.Send images in batches of 5. After each batch Claude asks questions. Answer short and honest: "I like the light", "the colors feel messy", "I don't know, I just feel it". That last one is a good answer too.
Here is batch [1] of images I LOVE.
For each image: name 3 concrete visual choices (light, color, composition, texture, styling, mood). No compliments.
Then: what do these images have in common? What is still unclear?
Ask me up to 3 questions.This is the step most people skip. Taste is defined just as much by what you reject. Then send the "almost" images.
Now a batch of images I DISLIKE.
For each one: what exactly makes it wrong for me? Compare it to the images I love.
Next I will send "almost" images: I like them, but something is off.
For each "almost" image, guess what is off. I will tell you if you are right.When Claude says it has enough, ask for the Taste Card. It's the passport of your taste.
Write my Taste Card based only on the images and answers I gave you.
Format:
1) The feeling I am after (2 lines)
2) Always: concrete rules (light, color, composition, texture, people, styling)
3) Never: concrete things to avoid
4) Tensions: the contrasts that make an image "mine" (for example: clean but raw)
5) Words to use in image prompts
6) Words to never use in image prompts
7) What you are still unsure about
Use concrete visual language. No generic words like "beautiful" or "aesthetic".Don't rush to build the skill. Test it first. Send your 6 test images without saying whether you like them. Claude has to guess.
Blind test. Here are 6 new images. I will not tell you what I think.
For each image, predict: LOVE / DISLIKE / ALMOST.
Give one reason per prediction, based on my Taste Card rules.
Then I will reveal my real answers, and you will update the Taste Card.Claude will get things wrong. That's normal, and it's exactly what you need. Every mistake is a new rule on the card.
You were wrong about image [number].
You said: [your prediction]. The truth: [my answer].
Why: [my reason, even if it is just a feeling].
Find the rule behind my answer and update the Taste Card. Show me only what changed.Leaving the card in one chat isn't enough: a new chat won't see it. So you turn it into a skill, an instruction file Claude switches on by itself every time you ask for an image or any visual.
Turn my final Taste Card into a Claude skill called mytaste.
Use your skill creator and follow its best practices.
The skill must:
1) Trigger every time I ask for an image, image prompt, video prompt, design, carousel or any visual, and every time I ask "does this match my taste?"
2) Contain my full Taste Card: feeling, Always, Never, Tensions, words to use, words to never use.
3) Before creating or prompting any visual, check it against the Taste Card and fix anything that breaks it.
4) When I send a new reference and say "update my taste", add the new rule to the skill.
Package it as a skill file I can save.Save the skill file Claude just built.
In Settings, open Capabilities, then Skills, and switch on mytaste.
In a new chat, ask for an image prompt. If you see "Using mytaste", the skill is working.
Now AI works through your eyes
The skill is on. Every visual request now runs through your taste automatically.
01A prompt for an image
Use my mytaste skill and write an image prompt for [Midjourney / ChatGPT images / other] for: [what you need].
Before writing, list which Always rules you will use.
After writing, check the prompt against my Never list and fix anything that breaks it.02A check for the result
Here is the image I generated. Score it against my mytaste skill from 1 to 10.
Name the 2 things that break my taste and rewrite the prompt to fix them.Every time you find something new you love or hate, send it to Claude and write: Update my taste. Your taste changes, so the skill should too.
Your evening plan: about 1 hour
Quick summary
- AI gives you the average. Taste is your job.
- Show, don't describe: what you love, what you don't, and the "almost" images.
- Claude finds the pattern. You check it with a blind test.
- Mistakes become corrections, and corrections become rules.
- Turn the card into a skill, and every new visual runs through your taste.
Your turn
Tonight, collect your first 20 images and open Claude. In an hour you'll have a system that works for you every day.
Want to learn how to create AI content in your own taste and earn from it? Vibe Directing. All courses: huva.io/courses.
More systems like this on @aiwithani on Instagram.
