Teach Your AI To Edit Like You (Codex/Claude Code System)

Start with a simple folder structure. You do not need a content platform or a database. You need a place where the agent can separate drafts, approved finals, rules, and reusable skills.

Use this structure:

email-copy-style-demo/
  emails/
    draft/
    final/
  rules/
    editorial-rules.md
  skills/

The draft/ folder is for work that is still changing. The final/ folder is for approved emails that can teach the system something. The rules/ folder is the shared memory. The skills/ folder is where the reusable instructions live.

You can make this more detailed later, but keep the first version boring. The point is to prove the loop before you design the perfect folder system.

Before you create a drafting skill, give the agent a starting style guide. The fastest way is to have it interview you.

Interview me about the editorial rules I use for marketing emails.

Ask about:
- audience
- tone
- subject lines
- claims or phrases to avoid
- CTA style
- examples I like

Turn my answers into rules/editorial-rules.md.

This is usually easier than writing the rules from scratch. Let the agent ask multiple-choice questions, answer quickly, and then clean up the result.

If you already have examples of good copy, add them too. Ask the agent to extract patterns from those examples and merge them into rules/editorial-rules.md. Good starter rules are specific enough to change the next draft:

• lead with the concrete workflow, not the tool category
• use short sentences
• avoid hype words like "revolutionary" or "game-changing"
• include one specific use case before making a broad claim
• keep CTAs direct and low-friction

Now create the first real test. Ask the agent to draft one email using the rules file, then save two copies: a working draft and a snapshot.

Draft a marketing email using rules/editorial-rules.md.

Save the working copy in emails/draft/.
Save an identical snapshot copy so we can compare it against my edited final later.

The snapshot is the important part. It should never change. It gives the agent a clean before-and-after comparison once you edit the email.

If you want the project to stay organized, have each email live in its own folder:

emails/
  draft/
    product-launch-email/
      prompt.md
      draft.md
      draft-snapshot.md
      final.md
      meta.md

The meta.md file can track the status:

Status: drafted
Rules Updated: no

That status becomes useful later when the automation needs to know which emails are ready to process.

Once the first draft-and-snapshot flow works, turn it into a reusable skill. This keeps you from rebuilding the same instructions every time you want a new email.

Use a prompt like this:

Create a skill where I can prompt with an idea for an email, promotion, or product launch.

The skill should:
1. Read rules/editorial-rules.md.
2. Create a new numbered folder for the email.
3. Save the assignment in prompt.md.
4. Draft the first version in draft.md.
5. Copy that exact first draft into draft-snapshot.md.
6. Create final.md for my edits.
7. Create meta.md with Status: drafted and Rules Updated: no.

Never overwrite draft-snapshot.md after it is created.
Never update editorial-rules.md from the first draft alone.

The skill should leave the email in a drafted state. Do not let the agent mark its own first draft as final. The whole system depends on the difference between the AI's first version and your edited version.

After the skill exists, test it with a small email assignment:

Use the email drafting skill to create a launch email for [PRODUCT_OR_WORKFLOW].

Audience:
[AUDIENCE]

Goal:
[WHAT_THE_EMAIL_SHOULD_GET_READERS_TO_DO]

CTA:
[CTA]

Check that the skill created both draft.md and draft-snapshot.md. They should match exactly at the start.

Now do the human part. Edit the working draft into the final approved version. Tighten the hook, remove weak claims, add concrete examples, and make the CTA sound like you.

When the edit is done, ask the agent to compare the snapshot against the final:

Compare draft-snapshot.md against final.md.

Identify the durable editorial lessons from my edits.
Update rules/editorial-rules.md with only the rules that will help future emails.

Do not add one-off campaign facts.
Do not duplicate existing rules.
Prefer fewer, sharper rules over a long list.

This is where the system starts learning. You are not describing your taste in the abstract. You are showing it the exact edits you made.

A useful rule update might say:

- When introducing a product, describe the specific use case before making a broad benefit claim.
- Avoid salesy reassurance phrases. Let the concrete example carry the promise.

A weak rule would say:

- Make the email better.
- Mention Purple Punch.

The first set changes future drafting behavior. The second set either says nothing or only applies to one campaign.

Once the two skills work, automate the review. The automation should scan for emails that are approved, compare the first draft with the final version, update the rulebook, and then mark the email as processed.

Use a prompt like this for the scheduled task:

Check the emails folder for the oldest draft that has been marked final, scheduled, or sent.

Use the editorial-rules-update skill on it.
Compare the original snapshot with the approved final.
Update rules/editorial-rules.md with durable lessons only.
Then update the email status so this draft does not run twice.

Daily scanning works well if you publish or approve email often. If you want a slower review cycle, run the deeper improvement process weekly. The key is that the automation should only process finished work. Do not let it learn from drafts that you have not edited yet.

Pro tip: Keep the rule update conservative. You want a rulebook, not a junk drawer. If the agent finds ten possible lessons, approve the two or three that would actually improve the next draft.