Start here · Lesson 1.3
A Week of AI Taste Training: An Illustrative Example
This is a worked example, and it is made up. No one ran this week, and none of the figures or results below are real. It is here to show how AI taste training could look across seven days for a small business, so you can copy the shape and measure your own numbers.
Say you run a small bakery and an AI helps you make a daily social post. Lesson 1.3 teaches the review method used here: Teach AI your taste.
The example shows a shape, not a result: collect a week, review it as a batch, save the notes as rules, and compare the next week.
The setup
Our imaginary bakery posts once a day. The owner has a tool that keeps standing instructions between sessions, like the CLAUDE.md files described in the Claude Code docs. Setup basics are in Lesson 1.1.
For seven days the owner does not give feedback. The AI makes a post each day and the owner simply saves them. This is the baseline.
Days 1 to 5: collect, do not fix
The owner resists editing and just writes notes in a text file as each post arrives. Imagine the notes look like this.
- Day 1: Text has a white outline. Looks cheap.
- Day 2: Same outline again. Photo is good.
- Day 3: A single word sits alone on the last line.
- Day 4: Text is pale grey on a light photo, hard to read.
- Day 5: Loaf is small in the corner. The bread should be the star.
Notice two things already. Notes are specific, and some repeat. That is why the review waits for a batch.
Days 6 and 7: the batch review
On day 7 the owner lays out all five or more posts and reads the notes together. The patterns are clear: outlines, orphan words, low contrast, small product.
The owner then says one sentence:
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 AI replies with the rules it saved. The owner opens the file and reads them.
What the rules might look like
| Note | Vague rule | Better rule |
|---|---|---|
| Outline | Make text look good | No white outlines on text |
| Orphan word | Fix line breaks | Never leave one word alone on its own line |
| Contrast | Readable text | Use a text colour that stands out from the photo |
| Small product | Show the bread | The loaf is always the hero of the image |
The right column is what you want. A stranger could follow it. The Claude Code docs say more specific and concise instructions are followed more consistently.
What could go wrong
An honest example includes the failures that are likely.
- A rule is ignored. The docs call these files context, not enforced settings, so the AI may still slip. Spot check each post.
- A rule is too broad. "Make it bold" could change things the owner wanted to keep. Tighten it with an example.
- Two rules clash. Find them when you read the file and rewrite one.
- The posts start to sound alike. Rules fix the look, not the voice. The owner adds a story or opinion each day.
What the owner would measure
No real numbers here, but these are the things worth tracking.
- Repeat notesCount how many notes in week two repeat week one.
- Hand editsCount how many posts needed changes before publishing.
- Rule hitsCheck whether each saved rule was followed in the new posts.
- Review timeNote how long the weekly review takes.
What to copy from this example
Copy the order: observe, batch, rule, check. Do not copy any expected result, because none was measured. Your numbers will depend on your business, your tool and how clear your notes are.
A second week, imagined
Say the owner repeats the process the next week. Here is what could be seen, again purely as an illustration.
The new notes might cover different problems, such as a caption that starts with the same phrase every day. A few old notes might return if a rule was too loose. The owner checks each saved rule against the seven new posts and marks followed or not followed. A rule that was missed gets shortened and given a concrete example, in line with the docs' advice that specific, concise instructions are followed more consistently.
None of this needs a promised outcome. The point is that the owner now has a record: which notes repeated, which rules held, which need work.
How to adapt the example to your business
Replace the bakery with your own thing and keep the structure.
- Service business. Your pieces may be emails or short videos. Notes could be about opening lines, length or calls to action.
- Online shop. Your pieces may be product descriptions. Notes could be about words to avoid and what to mention first.
- Solo creator. Your pieces may be scripts. Notes could cover pacing and phrases you never use.
If you have nothing to review yet, the lesson suggests starting from a brand you love, writing notes on its posts as if they were yours.
What this example cannot tell you
It cannot tell you how much time you will save, how much better your posts will be, or whether they will perform better with your audience. It is not evidence that the method works for any particular business. It is a worked sketch of the steps. Your own record of repeat notes, hand edits and rule hits is the only evidence worth trusting.
Questions people ask
Is this a real case study?
No. It is an invented example, clearly labelled, to show the method.
Can I run the same week?
Yes. Use your own posts, or a brand you admire if you have none yet.
When will I see a difference?
It varies. Compare week two with week one using the counts above rather than expecting a set outcome.
Where to go next
Run your own week with Teach AI your taste, Lesson 1.3, which reviews seven days of posts and saves every note as a rule. For measurements, read Measuring AI taste training.