The TEACH Framework for AI Taste Training

Doing AI taste training by feel gives patchy results: one week you review carefully, the next you skip it. A named routine fixes that. TEACH is a five step loop you can run every week, built on the idea in Lesson 1.3 that feedback disappears and rules stay: Teach AI your taste.

TEACH is a simple memory aid made up for this post, not an industry standard. Use it, change it, or ignore the name.

TEACH: Treat the week as a batch, Explain each note specifically, Add what you liked, Convert every note to a rule, Hold your own voice.

T: Treat the week as a batch

Collect seven days of output and look at it together. Put the pieces side by side in order, day one to day seven.

Why a batch? Single reviews feel quick but you end up repeating yourself. Seeing everything at once shows the notes that repeat, which are the ones worth turning into rules.

If you have no posts yet, use a brand you love as your starting set.

E: Explain each note specifically

Write notes that a stranger could act on. Use this test: could someone who has never seen your business follow the note and get it right?

• "Put text only on the left or at the bottom" passes.
• "Make the layout better" fails.
• "Never leave a single word alone on its own line" passes.
• "Fix the wording" fails.

The Claude prompting guidance says clear, explicit instructions help, and that applies to your notes as much as any prompt.

A: Add what you liked

Say which pieces were right, and why. "The product is always the hero" is a rule to keep, not just a ban. Without praise, the AI only knows the edges of what you do not want.

Official guidance on examples notes that a few relevant, varied examples improve consistency. Your approved pieces are your examples.

C: Convert every note into a rule

This is the step that makes the rest last. After your notes, say a sentence like this:

Apply this feedback to every post, not just the ones I mentioned. Then save each point as a rule, so you never make that mistake again.

Where the rules live matters. The Claude Code docs describe CLAUDE.md files you write and auto memory the tool writes, and both load at the start of each session. Setting that up is in Lesson 1.1.

H: Hold your own voice

Rules shape the draft. They do not provide your stories, your opinions, or the small true details that make a post yours. Add them yourself before you publish.

AI posting without your feedback is how work starts to sound generic. Holding your voice is the part you cannot hand off.

The routine on one page

1. TreatGather the week and view it as one batch.
2. ExplainOne specific note per problem.
3. AddName what worked and why.
4. ConvertAsk for each note to be saved as a rule.
5. HoldAdd your own voice, then publish.

A weekly rhythm

Pick a fixed slot, say the same afternoon each week, and run TEACH start to finish. Keep it short. The lesson reviews a week of posts in about five minutes, so a first pass need not be long.

After the first run, add a checking step: open the saved rules and read them. The docs describe the /memory command for viewing and editing what loads. Remove vague rules and split any that hold two ideas.

Signs the routine is working

Do not look for dramatic changes. Look for small signals.

• Fewer notes in this week's review repeat last week's.
• Your notes get shorter because more is already covered.
• You change less by hand before publishing.

If none of these show up after a few weeks, go back to E. Vague notes are the usual cause.

TEACH applied to a different job

To show that the routine is not tied to posts, here is an invented example with customer emails. Say you send a weekly update to your customers.

• Treat: gather the last seven updates.
• Explain: "Subject lines over six words get cut off on my phone. Shorten them." "The button text says Click here, make it say what happens next."
• Add: "The update from Tuesday opened with one plain sentence and I liked it. Open every one that way."
• Convert: ask the AI to apply all notes to every email and save each as a rule.
• Hold: add one true detail from your week before sending.

The shape is identical. Only the notes change.

How to adapt the letters

You can rename the steps if that helps you remember. Some people prefer a shorter loop of three parts: collect, write notes, save rules. The extra steps in TEACH are there for two jobs that people forget: saying what you liked, and keeping your voice. If you already do those, shorten it.

Where TEACH can fail

• You skip T and review daily. You will repeat notes and write rules in a hurry.
• You rush E. Vague notes make vague rules.
• You skip the check after C. You will not know if the rule was saved well.
• You skip H. The output may start to feel like everyone else's.

Each failure has the same remedy: go back to that step and do it properly once.

Questions people ask

Do I need to follow TEACH exactly?

No. The order is a guide. Skip a step if it has nothing to add that week.

How many rules is too many?

When you stop reading them. The docs note that large instruction files cause problems, so merge or delete.

Does TEACH work for emails and pages too?

Yes. It only needs a batch of output, specific notes, and somewhere to save rules.

Where to go next

Watch the full week review in Teach AI your taste, Lesson 1.3. For a copy and paste version of the notes and rules, see a free AI taste training template. To avoid common slips, read 5 AI taste training mistakes.