Ad angle testing

Ads are not shipped, they are tested. One ad is an opinion with a budget attached. This skill turns a product into a set of ads that disagree with each other usefully.

The core idea

Most people vary the wrong thing. They render one ad four times with different music and call it a test. That tests nothing, because the four ads make the same claim — they win or lose together.

WRONG axis                 RIGHT axis
────────────────────────────────────────────────
different music            different claim
different creator          different audience
different colour grade     different friction
same promise               same product

Vary the angle: who it is for, and what it promises them.

Method

1. Find three real angles

An angle is a who plus a friction plus a claim. Write them out before any tool call:

A  busy parents · no time to cook breakfast · one minute, one cup
B  gym people   · protein shakes taste chalky · actually drinkable
C  small kitchen owners · no counter space   · it is the size of a mug

Three angles for the same product. Each one excludes the others — that is what makes the test readable.

Where to find them: the product page's FAQ and its objection-handling tell you the real hesitations. The headline tells you the weakest angle, because it was written to satisfy everybody.

2. Write one hook per angle

Use the hook-writing skill. Different shapes, not rewordings — a reversal for one, a number for another, a confession for the third. Three rephrasings of one hook produce three ads that teach you nothing.

3. Quote the set, not the ad

quote_ad_video { "product_image_url": "…", "script": "<angle A>", "variations": 1 }

Each angle is its own call, because each has its own script. Multiply by three and state the total to the user before you spend anything.

Use the variations parameter within an angle when you want the same claim rendered more than once — different delivery of one script. Use separate calls when the claim itself changes. They are different tests:

variations: 3     one claim, three executions   → which execution
3 separate calls  three claims                  → which claim

Find the claim first. Execution only matters once something is working.

4. Render them as a batch

Submit all three, collect the run ids, then poll. Submitting before polling is considerably faster than one at a time.

create_ad_video { "product_image_url": "…", "script": "<A>", "look": "natural", "aspect_ratio": "9:16" }
create_ad_video { "product_image_url": "…", "script": "<B>", … }
create_ad_video { "product_image_url": "…", "script": "<C>", … }

Hold everything else constant — same product image, same creator, same look, same length. A test with two variables has no result.

5. Hand over as a set, labelled

Give the user the three MP4s with their angles written next to them. An unlabelled batch of three files is unusable a week later when the numbers come in and nobody remembers which was which.

ad-A-busy-parents.mp4   "one minute, one cup"
ad-B-gym-taste.mp4      "actually drinkable"
ad-C-small-kitchen.mp4  "the size of a mug"

Reading the result

Tell the user what to look at, because the platform will offer them twenty numbers:

• Hook rate (3-second views ÷ impressions) judges the first two seconds.

A low hook rate kills the angle, not the ad.

• Hold rate (watched to end) judges the body.
• Cost per result judges the offer, and is the slowest to read.

A losing angle with a strong hook rate is worth rewriting. A winning hook with a bad hold rate means the body did not pay the hook off.

What fails

• Varying execution before finding the claim. The most common waste.
• Three angles that are the same angle. "Fast", "quick" and "saves time"

are one angle.

• Changing the product image between variants. Now you are testing the

photo.

• Stopping at one round. The first set tells you which direction; the

second set goes deeper in that direction.

Spending rules

• Every call spends credits and create_ad_video needs a paid plan.
• Quote all three and state the total before the first render.
• Failed runs refund in full.