Rebuild a Claude skill by reverse engineering your best and worst examples

Ben explains the Reverse Engineer Anything idea and uses a three-step process on his 6 best and 6 worst newsletters to rebuild his newsletter skill instead of writing instructions by hand.

Ben AI · YouTube · published Oct 16, 2026 · Content & Creative

Rebuild a Claude skill by reverse engineering your best and worst examples — video by Ben AI▶ Watch on YouTube
Prompt
Here are my 6 best and 6 worst [newsletters]. Compare them, work out what separates the good from the bad, and write a Claude skill that reliably produces the good version.

Visual breakdown

Show Claude your six best and six worst newsletters instead of describing the process, let it find the winning patterns, then have it reverse engineer a reusable skill from them.

The problem: Skills built by describing every step take a long time to write and turn out worse than skills learned from great examples.

From examples to a tested skill

  1. Best and worst examplesSix top and six bottom newsletters, ranked by open and click rates.
  2. Compare and find patternsClaude compares structure, hook, CTAs, tone, length and subject lines, and says why each worked.
    Dropped: guesses: every finding is tagged seen, inferred or unknown
  3. You review the patternsCheck what Claude actually saw versus what it is guessing before building anything.
  4. Build the skill"Now build the skill": Claude reverse engineers the whole process with no step-by-step instructions.
  5. Test on real inputFeed it an idea or a past YouTube video and compare against your own examples.
  6. Reusable skillNewsletters in your voice that follow the winning patterns, every time you run it.
Your examplesManual reviewClaude (Opus 5.5)Result

What Claude is told to compare

  1. Structure and length: How the best issues are laid out and how long they run.
  2. Hook and subject line: Opening lines and subject-line patterns.
  3. CTAs and tone: Where the call to action sits and how it sounds.
  4. Seen / inferred / unknown: Tag every finding so Claude can't invent reasons.

By the numbers

6 + 6best and worst newsletters used as examples
3example reels used for the Shorts and Reels skill
3 stepsgather examples, find patterns, build and test

Do it yourself

Ben's three-step process works for anything with a clear finished output, like a client proposal.

  1. Gather examples: Collect your best and worst outputs with real performance data, or a creator's top posts or work you like if you have none.
  2. Find the winning patterns: Ask Claude to compare them, explain why each likely worked or didn't, and tag findings seen, inferred or unknown.
  3. Review the findings: Read the patterns, for example open rates were similar but click rates differed, so the body and CTA matter most.
  4. Build the skill: Tell Claude to build the skill by reverse engineering each winning pattern.
  5. Test it: Run it on a new idea or a past video and check the output against your examples.

His free skill packages all three steps and waits for your approval before building.

Lessons from the video

  • Show, don't tell: newer models learn style and detail from examples better than from descriptions.
  • Include bad examples too, so Claude can see what separates good from bad.
  • The seen / inferred / unknown tag stops Claude from inventing reasons.
  • The new newsletter skill wrote shorter issues, moved the CTA earlier and kept some info behind the link.
  • For tasks without a clear output, like email triage, do the task once with Claude or use record a skill instead.

Distilled from the video’s captions

How to try it

  1. Open Claude.
  2. Copy the prompt and fill in the [placeholders].
  3. Watch Ben AI's video for the full walkthrough and setup.

Related use cases

Share