Teach AI Your Taste: A Complete AI Taste Training Guide

You ask AI for a post, say what is wrong with it, and it fixes that one post. Next week it makes the same mistake again. AI taste training is the habit that stops this: you review a batch of work, and you turn every note into a rule that stays.

This guide walks the whole path in order, from nothing to a working set of rules. Lesson 1.3 does the same thing on camera with a full week of posts: Teach AI your taste.

Feedback disappears when the chat ends, so after each review ask your AI to save every note as a rule it will read next time.

Why feedback disappears

Each session in a tool like Claude Code starts with a fresh context window. The official docs say two things carry knowledge across sessions: a file you write called CLAUDE.md, and notes the tool writes itself, called auto memory. Anything you only said in chat is gone when the session ends.

That is the whole problem. Your notes were good. They just were never written anywhere the AI reads at the start of the next job.

The docs also say these files are treated as context, not as enforced settings. So a rule makes a mistake less likely, not impossible. Plan to check the work, and to tighten a rule when it gets ignored.

What you need before you start

You need three things.

• A week of output to review. Seven posts, seven emails, seven product descriptions, whatever you make often.
• An AI setup that keeps files or memory between sessions. Lesson 1.1 covers a setup like that: Lesson 1.1.
• Ten minutes with no interruptions.

No posts yet? Pick a brand you love and use its posts as your starting point. You are teaching taste, and taste can start from an example you admire.

The steps, in order

1. Lay out the whole weekPut all seven pieces side by side, day one to day seven, so you can see patterns instead of single mistakes.
2. Write one note per problemShort, specific, in your own voice. "No white outlines on the text, they look tacky" beats "make it nicer".
3. Add what you likeMark the pieces you would post as they are and say why. Good examples teach as much as complaints.
4. Spell out layout and wordingSay where text goes, what colour it should be, which words it must never use.
5. Say the rule sentenceAsk the AI to apply your notes to every piece, then save each point as a rule.
6. Check the saved rulesOpen the file or memory and read what it wrote. Fix anything vague.

Step 1 to 3: review the week as a batch

Reviewing one piece a day feels easier, but you repeat yourself. Reviewing seven at once shows you which note appears three times, and that is the one that deserves a rule.

Go through the pieces fast. Do not rewrite anything. For each one, write a plain note about what bothers you or what you like. Aim for notes a stranger could act on.

Compare these two notes on the same image post.

• Weak: "The text is hard to read."
• Specific: "Put the text in a colour that does not disappear into the photo, and never leave a single word alone on its own line."

The second one can be turned into a rule. The first one only tells the AI you are unhappy.

Step 4: say what you like, too

If you only list problems, the AI learns what to avoid and has to guess what to aim for. Tell it which pieces were right and what made them right. "The product is always the hero of the picture" is a rule about what to keep, and it is just as useful as a ban.

Step 5: the rule sentence

After your notes, say one sentence like this:

Here is my feedback on all seven days. Apply it to every post, not just the ones I mentioned. Then save each point as a rule, so you never make that mistake again.

The first half stops it from fixing only the pieces you named. The second half is what moves your notes out of the chat and into something that lasts.

Step 6: read what it saved

Do not trust that the rules were saved well. Open the file, or in Claude Code use the /memory command, which the docs describe for viewing and editing what is loaded. Read each rule and ask whether a new person could follow it.

The docs say the more specific and concise your instructions are, the more consistently Claude follows them. Cut anything long and vague. Split any rule that holds two ideas.

What to do next week

Next week, produce the new batch, then review it the same way. Most notes should now be new ones, because the old mistakes are covered. If an old mistake returns, the rule was probably too vague. Rewrite it with a concrete example.

Over a few weeks the rulebook grows, and you spend less time repeating yourself. Do not promise yourself a specific saving, because it depends on how much you make and how clear your notes are.

Questions people ask

How many rules should I keep?

Only as many as you will read. The docs note that very large instruction files are a problem, so merge duplicates and delete rules that no longer matter.

Can I do this without Claude Code?

The idea works with any assistant that can keep standing instructions. Without that, paste your rulebook at the start of each job.

What if I have no work to review yet?

Start with a brand you admire. Collect a few of its posts and write notes about what you would copy.

Will the AI always follow the rules?

Not always. Official docs describe these files as context, not enforced settings, so spot check the output.

Where to go next

Open Teach AI your taste, Lesson 1.3, and follow the week review with the full example on screen. If you want a clean setup first, start with Lesson 1.1. Then try the reusable routine in The TEACH framework for AI taste training.