5 AI Taste Training Mistakes and How to Fix Each

AI taste training goes wrong in predictable ways. The notes are fine, but they are vague, they are not saved, or they never get checked. Below are five mistakes that waste time, each with a fix. They follow from Lesson 1.3, which shows the right habit on a full week of posts: Teach AI your taste.

Most wasted time comes from notes that never become rules, so fix the saving step first and the wording second.

Mistake 1: giving feedback that only lives in the chat

You tell the AI what is wrong and it fixes the post. Next week the same mistake appears.

Why it costs time: each session starts with a fresh context window, according to the Claude Code docs. What you said in chat is gone unless it is written somewhere loaded at the start of the next session.

The fix. After your notes, say: apply this to every post, then save each point as a rule so you never make that mistake again. Then confirm the rule was written. See Lesson 1.1 for a setup that keeps files between sessions.

Mistake 2: vague notes

"Make it more on brand" gives the AI nothing to hold.

Why it costs time: the AI guesses, you disagree with the guess, and you write another vague note. The docs say specific, concise instructions are followed more consistently.

The fix. Use the stranger test. Could someone new follow this note and get it right? Rewrite until yes.

• Instead of "nicer text", write "text in a colour that stands out from the photo".
• Instead of "better layout", write "text only on the left or at the bottom".

Mistake 3: only saying what you dislike

A list of bans teaches the AI the edges, not the target.

Why it costs time: it avoids your complaints and invents new problems in the space left over.

The fix. Add what you like and why. "The product is always the hero" is a rule about what to keep. Official guidance on examples says a few relevant, varied ones improve consistency, so point to your best piece as the example.

Mistake 4: reviewing one piece at a time

Daily reviews feel safe, but you write the same note three times.

Why it costs time: you spot patterns late, and each rule gets written separately and in a hurry.

The fix. Review a week at once. Lay out seven pieces and read your notes together. The note that appears most is your first rule. The lesson reviews a week of posts in about five minutes.

Mistake 5: never reading the saved rules

You assume the AI wrote good rules. Sometimes it writes vague ones, or two that clash, or a very long file.

Why it costs time: a rule that does nothing looks the same as a rule that works, until the mistake returns.

The fix. Open the file. The docs describe the /memory command for viewing and editing what is loaded. Read each rule, split ones that hold two ideas, delete ones that are no longer true. Keep the file short, since the docs note that very large files cause trouble.

A quick self check

1. Saved?Is each of your notes in a file or memory the AI reads next session?
2. Specific?Would a stranger act on each note the same way you would?
3. Balanced?Do you have as many "keep doing" notes as "stop doing" notes?
4. Batched?Did you review the whole week together?
5. Read?Did you open the saved rules after the last review?

One more mistake worth naming: dropping your voice

It is not in the five, but it matters. Rules shape the draft, but posting AI work without your own feedback and details is how things start to sound generic. Add a real story, opinion or detail each time you publish.

How to find your own costly mistake

You may not do all five. Use this short method to find the one that costs you most.

Look back at the last two weeks of notes. Find any note you wrote more than once. For each one, ask why it came back.

• If you never saved it, that is Mistake 1.
• If it was too vague to act on, that is Mistake 2.
• If you only said what you disliked, that is Mistake 3.
• If you only noticed the repeat late, that is Mistake 4.
• If you saved it but it was ignored or garbled, that is Mistake 5.

Fix that one first. Fixing five things at once makes it hard to see what helped.

A before and after for one rule

Here is a single rule improved. The example is invented.

Before: "Make the posts look more professional."

Problems: it is vague, it names no action, and it may clash with other notes.

After: "Use one font. Put text only on the left or at the bottom. No white outlines on text."

Why it is better: it is three concrete instructions a stranger could follow, and you can check each one by looking at a post.

The official prompting guidance says clear, explicit instructions help, and the Claude Code docs make the same point about memory files. Rewriting a vague rule into a few concrete ones is the cheapest improvement available.

A habit that prevents most of these

End every review with the same short checklist: saved, specific, balanced, batched, read. It takes a minute and catches most of the problems above. Put it at the bottom of your notes file, so it is there when you finish.

Questions people ask

Which mistake should I fix first?

Mistake 1. If feedback is not saved, nothing else sticks.

How do I know a note is specific enough?

Use the stranger test: someone new could follow it.

What if a rule keeps being ignored?

Shorten it, add a concrete example, and check it does not clash with another rule.

Where to go next

Practise the fixes on a real week of posts in Teach AI your taste, Lesson 1.3. If you want a routine that builds the checks in, read The TEACH framework for AI taste training.