Meta Ad Library Analysis Skill

First use in three steps

1. Unzip and read SETUP.md. It explains the local commands, input options and output folder.
2. Run setup and add your own key locally. Setup creates a virtual environment, installs dependencies, creates .env from env.example and appends missing ignore rules to .gitignore. It preserves an existing .env and existing ignore rules.
3. Run a three-ad pilot, inspect it, then expand. Confirm capture coverage, Gemini access, media and classifications before using a larger sample.

Public collection needs Chrome. If you are using imported media, run the setup check with its media-only option. The key stays on your machine; do not add it to a shared ZIP, report or chat.

What you get from the analysis

• A small, closely reviewed set of prominent ads selected through the library’s impressions ordering.
• A separate sample of recent launches, with its period and actual coverage stated.
• Creative families supported by individual ad observations, with setting and emotion recorded separately.
• A local HTML report with video, actual timestamped screenshots, family comparisons and mechanic-setting and mechanic-emotion matrices.
• Markdown and per-ad CSV exports, plus the structured JSON input and local media in the output folder. Share the whole folder so its media links work.

Start with two different questions

For a dated view, start with the last seven days. Expand to fourteen or twenty-eight days when the volume is too small for the question. Keep the selected window visible. If time is very limited, a close deconstruction of the top ad can be more useful than shallow labels across hundreds.

Classify the argument, the setting and the emotion separately

First create a record for each reviewed creative: its opening, visible action, spoken content when available, on-screen text, product role, payoff and call to action. Only then group the records into families. Preserve missing or uncertain evidence instead of filling it in.

A language-learning ad set in a bakery and one set in a restaurant might share a structure: a communication failure, a concrete cost, help, hesitation and a successful retry. The locations differ. The persuasive mechanism may be the same.

Emotion is another dimension. Disappointment after a setback is different from embarrassment in public, even when both stories use a failed conversation. A useful matrix keeps these distinctions visible and links every classification to the ads behind it.

Review the video closely enough to support the claim

Native Gemini video and audio analysis is the primary review path. The media helper then extracts actual video frames at proposed beat timestamps using PyAV and decoded presentation times, so each screenshot can be checked against the source.

Gemini’s native video path samples visual frames at about one frame per second by default. Fast cuts, small overlays and uncertain dialogue can need a targeted recheck against the original. Do not treat a model-generated transcript or interpretation as verified evidence. See Google’s video-understanding documentation.

For an optional visual-only fallback, the helper can export frames across the full duration at two-second intervals and include the final frame. That fallback is not audio transcription. Check spoken claims against audio separately.

A brief you can adapt

Use competitor-ad-analysis for [advertiser or Meta Ad Library link].

Review the top 10 distinct creatives by verified impressions ordering and the ads launched in [date range]. Save the report and supporting evidence in [local output folder].

For each reviewed creative, record the narrative mechanism, setting, emotion and source. Show how several ads within a family share a structure. State the capture coverage and any inaccessible media. Do not infer spend or profitability from the public library.

Check the evidence before using the findings

• Verify the sort. A URL parameter alone does not prove the displayed ranking was retained. Ad longevity is not proof of meaningful spend or performance.
• Keep counts traceable. One creative may appear under several ad IDs. Preserve the mapping so repeated media does not masquerade as additional ideas.
• Check the whole story. A contact sheet can miss the failure, hesitation, product reveal or final payoff that changes the interpretation.
• Reconcile the families. Every selected record should be accounted for, including unavailable media and cases that do not fit confidently.

Download the skill ZIPBack to the skills library

Want help applying this to your own creative program? Talk to RocketShip HQ.