Start here · Lesson 1.3
AI Taste Training vs Fixing Posts by Hand: Which Wins?
Say you fix the same problem in your AI's posts every week. You can keep editing each post by hand, or you can spend a little time on AI taste training so the fix is saved as a rule. Which is worth your time depends on how often you repeat yourself.
This post compares the two honestly. It has no benchmark numbers, because none were run for it. You get a side by side table and a simple way to decide. Lesson 1.3 shows the rule approach in full: Teach AI your taste.
If you correct the same thing more than twice, a saved rule costs less than the third and fourth manual fix.
What each approach means
The manual way. You read each draft, change what you dislike, and publish. The AI never learns anything. Every new draft starts from the same blank place.
Taste training. You review a batch, write specific notes, and ask the AI to save each as a rule. The next draft starts with those rules loaded. Official Claude Code docs say each session starts with a fresh context window, and that CLAUDE.md files and auto memory are what carry knowledge forward.
Side by side
| Question | Manual fixes | Saved rules |
|---|---|---|
| Setup effort | None | Some, once |
| Cost per repeat mistake | Full fix every time | Written once |
| Control | Total, per piece | Rules guide, you still check |
| Risk | You get tired and skip fixes | Vague rules get ignored |
| Best for | One off work | Work you repeat weekly |
| Consistency | Depends on your mood | Better, not guaranteed |
Where the manual way wins
Manual editing is the right call in a few cases.
- You make something once and never again.
- The fix is a matter of facts, such as a wrong date or price.
- The change depends on a single piece and would not apply elsewhere.
Do not turn a one off correction into a rule. A rule file full of exceptions gets long and confusing, and the docs warn that very large instruction files cause trouble.
Where saved rules win
Rules win when the problem is about style and returns. Colour of text, words you never use, where the product sits in the frame, how a caption opens. These are the notes that appear in every review.
Reviewing a week at once also helps you notice them. A note you write three times is a rule you have not saved yet.
The honest limits
Rules are not magic. The Claude Code docs say the memory files are treated as context, not enforced configuration, and that more specific, concise instructions are followed more consistently. So expect fewer repeats, not zero. You still read the output.
There is also the setup cost. You need a tool that keeps files or memory between sessions, covered in Lesson 1.1, and a few minutes to write rules well.
A simple way to decide
Use this checklist over your last two weeks of work.
- List your editsWrite down every change you made by hand, one line each.
- Mark the repeatsCircle any edit that appears in more than one piece.
- Count themIf you have three or more repeated edits, they are candidates for rules.
- Write them as rulesOne sentence each, specific enough for a stranger.
- Review again next weekSee whether the same edits come back.
If almost nothing repeats, keep editing by hand and skip the setup.
A worked comparison, illustrative only
Say you run a small bakery account and the AI keeps putting text in a thin white outline. By hand, you remove the outline from each image, each week. With a rule, you write "no white outlines on text, they look tacky" once and check that the next batch obeys. Whether the rule saves you time depends on whether it is followed, which you can only see by looking. This is an invented example to show the logic, not a measured result.
A middle path: rules for style, hands for facts
Most people end up using both. Split your edits into two piles.
Style edits are about how things look and sound: colours, layout, banned words, how captions open. These repeat, so they are rule candidates.
Fact edits are about what is true this week: a date, a price, a stock level, an opening hour. These change every time, so they stay manual. Putting them in a rule file only adds clutter.
When you are unsure which pile an edit belongs in, ask whether you would make the same change next month. If yes, it is probably style.
What the setup actually involves
People overestimate the setup. For taste training you need somewhere the AI reads at the start of each session, and a habit of writing notes that a stranger could follow. In Claude Code, that can be a CLAUDE.md file you write, or auto memory that the tool writes from your corrections, both described in the official docs. You then spend a few minutes a week on review.
The part that takes real effort is wording. Vague notes make rules that do nothing, so the time you save depends on how well you write them. That is a skill, and it improves with repetition.
When the manual way quietly costs more
Manual editing feels free because the cost is spread out. A two minute fix per post does not look like much until you do it every day for a month. Add the small annoyance of remembering what you changed last time, and the habit starts to wear on you.
There is also a risk of skipping edits when you are busy, which makes your output uneven. A saved rule is not tired. It is also not perfect, so a quick check stays in the plan.
Questions people ask
Is taste training only for coders?
No. Claude Code is a coding tool, but the memory idea works for anyone producing posts, emails or pages.
Can I use both approaches?
Yes, and most people will. Rules cover style, hand edits cover facts and one off changes.
How do I know the rules are working?
Count how many notes in each weekly review repeat last week's notes.
Where to go next
If you edit similar things each week, try Teach AI your taste, Lesson 1.3, and see the rule approach on a full week of posts. For the first setup, see Lesson 1.1. For a repeatable routine, read The TEACH framework for AI taste training.