mcpField
A
ad-library-teardown

ad-library-teardown

github.com/scrapecreatorsMarketing & growth5.0k1.4k installsUpdated 9w ago
Repository

About ad-library-teardown

A marketing skill for analyzing public ads in Meta, Google, and LinkedIn libraries. It organizes competitor messaging, offers, and creative angles into practical findings and test ideas.

Key features

  1. 01Search ads and advertisers
  2. 02Extract copy, CTAs, and offers
  3. 03Analyze video ad transcripts
  4. 04Group ads by messaging angle
  5. 05Suggest tests from repeated patterns

Use cases

  1. 01Compare competitor ad messaging
  2. 02Collect hooks and landing-page angles
  3. 03Extract messaging from video ads
  4. 04Plan tests based on competitor ads

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/scrapecreators/social-media-research-skills --skill ad-library-teardown

After installing, type/ad-library-teardownin Claude Code to confirm it loaded.

Details

From the project README

Ad Library Teardown

Overview

Analyze public ads to understand a competitor's messaging, offers, creative strategy, and testing angles. The output should be a practical teardown marketers can use to write better ads or decide what to test.

When to Use

Use this skill when the user asks to:

  • analyze a competitor's active ads
  • search Meta/Facebook, Google, or LinkedIn ad libraries
  • extract ad hooks, CTAs, claims, offers, and landing page angles
  • compare ad messaging across competitors
  • summarize video ad transcripts
  • generate ad test ideas from competitor ads

Data Sources

Ad librarySearch/list endpointDetail endpointTranscript endpoint
Meta/Facebook/v1/facebook/adLibrary/search/ads, /v1/facebook/adLibrary/company/ads, /v1/facebook/adLibrary/search/companies/v1/facebook/adLibrary/ad/v1/facebook/adLibrary/ad/transcript
Google/v1/google/adLibrary/advertisers/search, /v1/google/company/ads/v1/google/adn/a
LinkedIn/v1/linkedin/ads/search/v1/linkedin/adn/a

Workflow

  1. Find the advertiser

    • Use company search endpoints when the user provides only a brand name.
    • Use domain/advertiser/page IDs when available.
  2. Fetch active ads

    • Prefer active ads unless the user asks for historical analysis.
    • Capture platform, advertiser/page, ad ID, start date, creative type, text, headline, CTA, destination URL, and source URL.
  3. Fetch details for representative ads

    • Enrich the ads with detail endpoints.
    • For video Meta ads, fetch transcripts when available.
  4. Cluster messaging Group ads by:

    • pain point
    • persona
    • offer
    • proof/social proof
    • feature/benefit
    • objection handled
    • comparison/alternative angle
    • urgency/discount
  5. Extract swipeable elements

    • hooks
    • headlines
    • primary text patterns
    • CTAs
    • claims
    • offers
    • visual/creative concepts
  6. Recommend tests Suggest tests based on repeated patterns and gaps, not random ideas.

Output Format

# Ad Library Teardown: {brand}

## Summary
- Ads analyzed: {count}
- Platforms: Meta / Google / LinkedIn
- Main positioning:
- Strongest repeated offer:

## Messaging Angles
| Angle | Evidence | Example ads | Notes |
|---|---|---|---|

## Hooks and Headlines Swipe File
- "..."
- "..."

## Offers and CTAs
| Offer | CTA | Platform | Example |
|---|---|---|---|

## Video Transcript Notes
- [Ad](url): summary, hook, best quote

## What They Appear to Be Testing
1. ...
2. ...

## Recommended Tests for Us
1. ...
2. ...
3. ...

## Sources
- [Ad](url)

Common Pitfalls

  • Do not claim an ad is winning just because it is active. Say it is active or repeated; performance is not public unless the endpoint returns it.
  • Do not ignore repeated ads. Repetition is often a useful signal.
  • Do not invent spend, conversion rate, or targeting unless public data includes it.
  • Do not skip video transcripts when the user asks for hooks or messaging from video ads.

Comments

Markdown supported