Start here · Lesson 1.3
10 AI Taste Training Strategies Worth Trying
Most advice about teaching AI your style is a vague tip to "give it examples". This list is more specific. It gives you ten AI taste training tactics you can try, with a note on when each one fits. None of them comes with a promised result. Pick two, run them for a week, and keep what helps.
The core idea comes from Lesson 1.3: feedback disappears, rules stay. You can see it done with a full week of posts in Teach AI your taste.
Pick two tactics, use them for one week, then keep only the ones that made your next review shorter.
1. Review a whole week at once
Look at seven pieces together instead of one a day. Patterns show up, and a note that appears three times is a rule waiting to be written.
2. Say "make that a rule" after every note
Say it as often as it takes. The sentence moves a comment out of the chat and into something the AI reads next time. Without it, the note only lives in one session.
3. Make every note specific
Write notes a stranger could act on. "No word left alone on its own line" is usable. "Looks off" is not. The official prompting guidance says clear, explicit instructions work better, and that holds for rules too.
4. Praise out loud
Say what you liked and why. "The jar is always the hero" tells the AI what to keep. A list of bans only tells it what to avoid.
5. Spell out layout
Layout is easy to get wrong and easy to fix with words. Say where text goes, such as on the left or at the bottom, and where it must not go.
6. Save rules where the AI reads them
In Claude Code, the docs describe CLAUDE.md files you write and auto memory the tool writes. Either way, the point is the same: store the rule somewhere loaded at the start of each session. Setup is covered in Lesson 1.1.
7. Read the saved rules
Open the file or use the /memory command described in the docs. Delete vague rules and split rules that hold two ideas. Concise, specific instructions are followed more consistently, according to the docs.
8. Add an example to each rule
Official guidance on examples says a few well made ones improve accuracy and consistency, as long as they are relevant and varied. Pair a rule with one good example and, if useful, one bad one.
9. Start from a brand you love
No posts yet? Pick a brand whose posts you admire and write notes about them. You can teach taste before you have output of your own.
10. Keep adding your own voice
Add your stories, opinions and details after the AI drafts. Rules shape the draft, but only you can supply what is true about your business.
| Tactic | Best when | Effort |
|---|---|---|
| Weekly batch review | You publish often | Low |
| Make that a rule | Every time | Very low |
| Specific notes | Notes keep coming back | Low |
| Praise and examples | Output is average, not wrong | Medium |
| Reading saved rules | Rules pile up | Low |
How to choose your two
Look at what annoys you most this week.
- If the same mistake keeps returning, start with 2 and 6.
- If your notes feel vague, start with 3 and 5.
- If the output is fine but bland, start with 4 and 10.
- If you have nothing to review, start with 9.
A one week test
Run your two tactics for a week, then count something simple: how many notes in this week's review were repeats from last week. Fewer repeats is a good sign. If the number does not move, the rules are probably too vague, so rewrite them.
Three ways to combine them
Tactics work better in pairs. Here are three pairings that fit common situations.
The beginner pair: batch review plus the rule sentence. Look at the week together, then say "make that a rule" after your notes. This needs no extra tools beyond somewhere to save rules, and it covers the core of the lesson.
The tidy pair: specific notes plus reading saved rules. If your rule file already exists but feels messy, rewrite notes using the stranger test, then open the file and clean it. The Claude Code docs describe the /memory command for this. Concise, specific entries are followed more consistently, so this pair pays off when the file has grown.
The voice pair: praise plus your own details. If the output is correct but bland, say what you liked and add a real story or detail each time you publish. Rules cannot supply what only you know.
Mistakes to avoid when trying these
- Trying all ten in week one. You will not know which did anything.
- Writing rules for one off facts. A wrong price is a fix, not a rule.
- Never pruning. The docs warn that very large instruction files cause problems.
- Treating rules as locks. They are context, so keep checking the work.
What to write down after the test week
Keep three lines in a note: which two tactics you tried, how many notes repeated from the week before, and one thing you would change. That small record tells you whether to keep going, swap a tactic, or rewrite your rules. You do not need a dashboard for this.
Questions people ask
Do I need all ten?
No. Two done well beat ten done halfway.
What if the AI still ignores a rule?
The docs describe these files as context, not enforced settings. Make the rule shorter and more concrete, and check the output.
How long should a review take?
The lesson reviews a week of posts in about five minutes. Yours may take longer at first.
Where to go next
Lesson 1.3 shows the first six tactics in one sitting: Teach AI your taste. For a routine you can repeat weekly, read The TEACH framework for AI taste training. For the errors to avoid, see 5 AI taste training mistakes.