16 Claude Skills for Ad Creative

The problem this pack exists for

An ad that ran well for six weeks starts sliding. Click-through rate is down by about a quarter. The ad has been live long enough that everyone already has a word for what is happening, and the word is fatigue.

Three different things produce that exact shape on a chart, and only one of them is the creative.

The creative genuinely wore out. The same people saw it too many times and stopped reacting. This is real, it happens, and it is the case everyone assumes. Its signature is a click-through rate falling while frequency climbs, both in the same window.

The audience got smaller. Delivery narrowed onto a shrinking pool. Frequency climbs, click-through rate holds. The finding here is the audience, not the ad, and replacing the creative changes nothing except which ad is stuck in the same small pool.

The auction got more expensive. Cost per thousand impressions rises, the campaign buys different, worse-matched impressions to hold its budget, and click-through rate falls while frequency stays flat. Nothing about the ad changed. A competitor, a season or a bid strategy did.

The three look identical if you only look at click-through rate, which is what almost everybody does, because it is the column that moved. The cost of getting it wrong is not symmetrical either. Retiring a working creative feels like decisive action and produces a visible improvement for about a week, because the replacement launches into the cheaper part of the auction. The account then learns a lesson that is not true and repeats it.

That asymmetry is the whole reason this pack exists. Creative work is full of decisions that are cheap to make, expensive to unmake, and almost never checked afterwards.

What is in the pack

Sixteen skills, in four groups. Each one is a markdown file with required inputs, a workflow, decision rules with thresholds, an output format, a worked example and guardrails.

Writing

• Search headline set. Builds a full Responsive Search Ad asset set for one ad group, decides what gets pinned, and names which headlines are the same headline twice.
• Paid social primary text. Meta and LinkedIn primary text, headline and link description, built around the point where a feed reader stops reading.
• Video hook lines. The first three seconds as three synchronised parts: the spoken line, the on-screen line and the opening shot.
• Funnel stage variants. Takes one angle that already works and rewrites it for cold, warm and retargeting, naming what changes and what must stay identical.
• Offer and proof line. Turns a vague benefit into a stated offer, the proof that carries it and a risk reducer. Refuses to supply proof you have not evidenced.

Preparing

• Creative brief from account data. Builds the next production round out of what already ran, so the brief comes from evidence rather than a brainstorm.
• Winning ad teardown. Decomposes one ad into angle, promise, proof and structure, and separates what transfers from what does not.
• Angle inventory. Counts the arguments the account has genuinely tested, once the duplicates collapse into each other.

Checking, before it runs

• Claim safety review. Reads finished copy line by line for unsourced numbers, unqualified superlatives and restricted-category wording.
• Ad to landing page match. Checks the destination delivers the promise, the offer and the call to action where an arriving reader can see them.
• Asset coverage check. Counts distinct angles, formats and aspect ratios live in an ad set and says whether the account can learn anything from a week of spend.

Reading what happened

• Early signal cut. Most new creatives will never reach a readable conversion sample. This one decides which to cut in their first days on leading indicators, and states the error rate it accepts for doing so instead of hiding it.
• Creative fatigue read. The problem above, with a rule instead of a feeling.
• Creative test readout. Says whether a variant won or the result is a tie, and reports ties as ties.
• Losing ad diagnosis. Locates the stage where an ad loses people, so a hook problem, an offer problem and a delivery problem are told apart rather than guessed at.
• Weekly creative readout. What to ship, cut and test this week, in language a client reads without a glossary.

The three rules that do the most work

Fatigue needs two signals, never one. The pack will not return a fatigue verdict on a falling click-through rate alone. It wants a second signal in the same window: frequency rising against its own prior level, the creative reaching fewer new people for the same money, or the absolute frequency ceiling crossed. If only the click-through flag fired, it says so and routes you to the audience or the auction instead. It also refuses to read frequency as a lifetime figure, because lifetime frequency only ever rises and will eventually prove fatigue on any ad old enough.

A winner has to clear two gates, and they fail in opposite directions. One asks whether a difference is big enough to be worth acting on. The other asks whether it is there at all. A fixed percentage gate on its own hands you a confident winner on fifty conversions a side, where the gap is inside noise; a significance test on its own tells a high-volume account to act on a difference too small to pay for the work. The readout applies both and tells you which one failed, because those are different next moves: run it longer, or leave it alone.

Unequal delivery voids a comparison outright. Paid social does not split impressions evenly between the ads in an ad set and does not go back and rebalance. A variant that took a tenth of the test ran against different people at a different time, so the difference you are reading is the allocation, not the creative. That is confounding rather than noise, and no amount of extra sample repairs it. The readout says so before it reports any gap.

All three exist because the failure they prevent is invisible. Nobody ever finds out that the ad they retired would have kept working, or that the hook they rolled out across the account won by noise.

What the skills refuse to do

Six of the sixteen decline to answer under stated conditions rather than return a number the data cannot carry.

The test readout refuses on a variant under the volume floor, even when its rate looks level with the control, because a rate on a thin sample is a coincidence with a decimal point. It also throws the whole comparison out when one variant took a tenth of the impressions, because that is a confound rather than noise and more data does not fix a confound. The fatigue read refuses on one signal, and refuses again if an automated creative enhancement was switched on between the two windows, because then the decline belongs to a different asset. The teardown refuses to conclude anything about a competitor's results, since a public ad library shows that an ad ran and nothing else, however long it ran for. The angle inventory refuses to call an angle tested on an ad that never got spend. The landing page check refuses to pass a promise it could only find below the fold. And the claim review refuses to soften an unevidenced claim into a vaguer one, because a vague unevidenced claim is the same defect with worse copy.

That is the part worth taking even if you never install any of this. A check that always returns something is not a check. Most creative reporting failures are not wrong numbers; they are real numbers answering a question nobody asked.

How the numbers work

Every threshold in the pack lives in one file and is cited by name from the skills, so a number cannot drift between them. The helper scripts read their values out of that same file at run time rather than carrying their own copies, which is the only version of this claim that survives someone changing a number. The pack ships a validator that fails the build if a skill restates a threshold instead of citing it, cites a key that does not exist, or writes a worked example that never declines anything, and a self-test that recomputes the worked example by an independent route and asserts the scripts agree with it.

Each threshold carries a tag. Platform guidance means Google, Meta or LinkedIn state it themselves. Heuristic means a practitioner starting point to recalibrate against your own account. Most of this pack is heuristic and says so. LinkedIn's truncation point, for instance, is tagged as a heuristic rather than platform guidance, because LinkedIn does not publish it and it moves by device.

How to install

The pack is on GitHub under an MIT licence. Clone it and copy the skill folders into your Claude skills directory, either inside a project or at user level.

git clone https://github.com/mardab96/ad-creative-claude-skills.git
mkdir -p ~/.claude/skills
cp -r ad-creative-claude-skills/*-ad-creative \
      ad-creative-claude-skills/references \
      ad-creative-claude-skills/scripts \
      ad-creative-claude-skills/examples \
      ad-creative-claude-skills/AGENTS.md \
      ~/.claude/skills/

The references folder is not optional: it holds every number the skills cite, and without it the thresholds do not resolve. Start a new session in that project and the skills activate when what you are asking about matches one of them. If you use Claude somewhere other than Claude Code, the files are plain markdown and you can paste the relevant one into the conversation with your export underneath.

Two skills ship a small Python helper for arithmetic that should not be estimated: the windowed fatigue signals, and the sample-size and gap check on a finished test. The install command above copies them.

Where to start

If you ship a lot of creative and most of it never reaches a readable sample, start at Early Signal Cut. It is the one skill here built to give you a decision rather than an honest refusal, and it tells you what that costs.

Open Angle Inventory first on any account you have inherited. It is the cheapest skill in the pack and it usually resets the question, because an account that has run two hundred ads has often argued three things.

If something is already going wrong, start at Creative Fatigue Read rather than at a rewrite. The most expensive routine mistake in creative work is replacing an ad that never stopped working, and it is the one nobody ever gets told about.