Every time you fix an AI draft and move on, you throw away the most valuable data you have. The edit itself is a lesson. The Red-Pen workflow captures those lessons, tags them, and converts them into prompt instructions, so your AI gets better every week instead of making the same mistake forever. It's the working version of the Feedback Loops pillar.
The five tags
Every edit you make to a draft gets one tag. Five is the whole system; resist inventing more.
- TONE: the words were fine but the feeling was wrong. Too stiff, too hype-y, too apologetic.
- CLARITY: a reader would stumble. Long sentence, buried point, jargon, vague claim.
- CTA: the ask was weak, generic, missing, or in the wrong place.
- HUMAN: it reads like AI wrote it. Stock phrases, symmetrical sentence stacks, summary closers, fake enthusiasm.
- LEGAL: can't say that. Unverifiable claims, invented stats, promises we don't make, compliance issues.
Step 1: Edit with tags (10 minutes)
Take your next AI draft. Edit it the way you normally would, but log each change as you go:
| # | Original | Your edit | Tag |
|---|---|---|---|
| 1 | |||
| 2 | |||
| 3 | |||
| 4 | |||
| 5 | |||
| 6 | |||
| 7 | |||
| 8 | |||
| 9 | |||
| 10 |
Ten edits is usually one or two drafts' worth. Small edits count; "changed 'utilize' to 'use'" is a perfectly good row.
Step 2: Find the patterns (5 minutes)
Sort your log by tag. Look for repeats. One TONE edit is a fluke; four TONE edits that all soften hype are a rule waiting to be written.
Ask of each cluster: what single instruction, given before the draft was written, would have prevented all of these edits?
Step 3: Convert to prompt instructions (10 minutes)
Turn your top three clusters into instructions and add them to your standing prompt. Write them as rules with examples, because AI follows examples better than adjectives:
Rule format: [Do/Don't] + [specific behavior] + [example]
Add your three:
Keep a running "prompt rules" doc. Every Red-Pen session adds up to three rules. When the doc passes ten rules, prune: delete any rule the AI no longer breaks.
Worked example
A consultant ran Red-Pen on two AI-drafted newsletter sections. Her log:
| # | Original | Her edit | Tag |
|---|---|---|---|
| 1 | "In today's rapidly evolving business landscape" | Deleted, started with the client story | HUMAN |
| 2 | "utilize proven methodologies" | "use a method that works" | TONE |
| 3 | "We are thrilled to announce" | "Quick news:" | HUMAN |
| 4 | "significantly improve outcomes" | "cut onboarding time from 3 weeks to 1" | CLARITY |
| 5 | "Don't hesitate to reach out!" | "Reply to this email and tell me your version." | CTA |
| 6 | "empower your team to succeed" | "give your team a checklist they'll actually use" | TONE |
| 7 | "It's important to note that" | Deleted | HUMAN |
| 8 | 62-word sentence about scheduling | Split into three sentences | CLARITY |
| 9 | "guaranteed results" | "here's what past clients saw" | LEGAL |
| 10 | Closing paragraph restating the whole email | Deleted, ended on the CTA | HUMAN |
Her clusters: HUMAN (4 edits), TONE (2), CLARITY (2). Her three new prompt rules:
- "Never open with a scene-setting phrase about landscapes, eras, or 'today's world.' Open with a specific person, moment, or number. Example: 'Last Tuesday a client asked me...'"
- "Ban these words: utilize, empower, thrilled, hesitate, note that. Use the plain verb instead: use, give, glad, ask."
- "Never end by summarizing. End on the call to action. The last sentence is always an ask or a question."
Next draft, she made four edits instead of ten. That's the loop working.
Make it stick
- Run Red-Pen once a week on whatever you edited anyway. It adds 10 minutes to work you were already doing.
- Date your prompt rules. When a rule is 3 months old and the AI never breaks it, retire it to keep the prompt lean.
- If two people edit content, compare tags monthly. Where your tags disagree (one person's TONE edit is another's non-issue), you've found a voice decision nobody made yet. Make it, write it down, move on.