Winning Ad Teardown Ad Creative

Winning Ad Teardown

Shared quality bar: ../references/output-standard.md. All numbers cited here live in ../references/thresholds.md. The output is a pattern with an explicit transferability verdict, not praise for an ad somebody already likes.

Use this skill when

One specific ad has been singled out and the next question is what to copy from it. Typical triggers: a competitor ad someone screenshotted from a public ad library, an ad a client keeps citing, or your own single best performer that the team wants to clone. The unit of work is one ad. If the question spans the whole library, use angle-inventory-ad-creative. If the deliverable is a set of production slots, use creative-brief-from-account-data-ad-creative.

Required input

• The ad itself: image or video, headline, primary text, call to action button, and the destination URL.
• Its source: own account, or a public ad library with the library page cited.
• For an own ad: spend, impressions, link CTR, conversions and the live dates, so the claim that it won can be tested.
• For a competitor ad: first seen and last seen dates only, nothing else is available.
• The audience and offer the ad ran against, where known.

Analysis workflow

1. Establish provenance before analysis. Own ad with an export, or competitor ad from a library. The two paths produce different evidence labels and different ceilings on confidence, and mixing them is the main failure of this exercise.
2. For an own ad, test whether it actually won: check the window against test.min_runtime, the volume against test.volume_floor, and the gap to the next ad against the right relevance key: money.relevance_floor where the gap is in cost per acquisition, which owns that comparison everywhere in this pack, and test.material_gap where it is in a rate such as link CTR. Say which you used. An ad inside whichever key applied is a tie and is torn down as a sample, not as a winner.
3. For a competitor ad, check it clears teardown.observation_floor before spending any effort on it, then apply reporting.no_competitor_spend in the first line of the output. Longevity shows a decision somebody kept making. It shows no spend, no result and no verdict, and any inference from run length is labelled hypothesis.
4. Set the ceiling now, before decomposing anything, because this is where the exercise usually goes wrong: it neuters the evidence correctly in one sentence and then carries on producing recommendations as though it had not. Where the ad sits below teardown.observation_floor, or where the only evidence available is a public ad library whatever the run length, this skill produces exactly one thing: the structure. It states plainly that structure is the only transferable output, that no performance conclusion is available from this input, and that nothing here says the ad worked. It then writes the structure with no recommendation attached to it: no slate, no ranking, no "worth testing", because a recommendation is a performance claim wearing a verb. What the reader does with the structure is their decision, made on their own evidence, and the output says so in that many words.
5. Decompose the ad into four named parts: the angle as a sentence, the promise made to the reader, the proof offered for that promise, and the structure of the opening, which is the window named in limits.video_hook_seconds for video and the first line for feed text, read against limits.meta_primary_text_visible.
6. Classify every element as transferable or not. Transferable means it survives without this brand's price point, distribution, catalogue, recognition or licensed footage. Name the specific dependency for each element you rule out.
7. Check the destination. An ad whose promise is carried by its landing page is only half copyable from the ad alone, and the page check belongs to ad-to-landing-page-match-ad-creative.
8. Convert what survives into one testable variant, and only where the ceiling in step 4 did not fire. On the library-only path this step does not run: there is no variant to propose, because proposing one asserts the pattern is worth spending against and nothing in the input supports that. Otherwise, one variant against a named control, one changed element only (test.one_change_rule), and mark any factual line the pattern would require with [claim: needs source] under claims.evidence_rule.
9. State what you declined to lift and why, including anything touching claims.restricted_topics.

Decision rules

• Own ad clearing volume, runtime and gap: ship the pattern into the next round, confidence high.
• Own ad under test.volume_floor: monitor and name a read date. Confidence is low whatever else is in place.
• Competitor ad, or any ad below teardown.observation_floor, or any ad whose only evidence is a public ad library: the output is structure only. No performance conclusion, and no recommendation attached to the structure. The decision column on every row reads needs_data, not test, because test is a recommendation and this input cannot support one. Say in the opening line that the pattern is offered as description and that the decision to try it belongs to whoever holds evidence this skill was not given.
• Element depends on brand recognition, catalogue depth or a licensed asset: cut it from the pattern and say which dependency.
• Element is a format or pacing choice with no brand dependency: test.
• Any lifted line stating a number or an outcome: approval_needed until the user supplies the source.

Output format

Open with one line naming what this ad is evidence of, and what it is not.

Then the pattern written as one paragraph a writer can build from, then the single variant to test, then What this could not see, Missing data, Approval gates.

On the structure-only path, which is every competitor ad and every ad below teardown.observation_floor, the variant block is absent rather than empty and the output says why in one line: no performance conclusion is available from this input, so the structure is offered as description and carries no recommendation. An output that ends in something to try has quietly promoted longevity into evidence, which is the failure this whole skill exists to prevent.

Practical example

Illustrative made-up data, not a real advertiser. A competitor ad has been live in a public library for 214 days across 6 creative variants. The team wants it called a proven winner.

Reading: refused. Under reporting.no_competitor_spend, 214 days shows only that somebody kept the ad on. Spend is needs_data, results are needs_data, and the opening line says so. What can be read is structure: the first line is 19 characters and lands inside limits.meta_primary_text_visible, the proof is a named third party rather than a self-reported figure, and the angle is a single sentence about switching cost.

Transferable: the switching-cost angle and the short opening line. Not transferable: the named third party, which the account does not have, and a "rated number one" line, cut under claims.superlative_rule because no comparison was supplied.

Output: the structure, and nothing else. The 214 days clear teardown.observation_floor, so the decomposition was worth doing at all, but the only evidence is a public ad library, so no variant is proposed and no row carries test. Every performance row is needs_data. The team asked for a slate built on this ad and does not get one: what they get is a description of how the ad is built, with the decision to spend against it left to whoever holds account evidence this teardown was never given.

Guardrails

• Never treat run length, comment count or reaction count as performance.
• Never lift a claim, statistic, customer name or review from another brand's ad.
• Never call an own ad a winner inside money.relevance_floor on cost per acquisition, or inside test.material_gap on a rate.
• Do not fill a missing dependency with an invented equivalent for this brand.