A skill that makes Claude work as the creative lead on your paid social and search ads, one stage at a time, with you choosing at each fork. It turns your notes into an ad brief and a proof bank that marks every fact as confirmed, needing evidence or unusable. It writes five angles, scores them and helps you pick three. Then it lays out one static concept per angle for every shape you buy, writes image prompts and assembly specs that keep your product exact, writes the copy for Meta, TikTok, Pinterest and Google within their limits, plans a small test and checks every ad before it runs.
Inside are four reference files. One holds the platform specs: sizes, safe zones in pixels, type sizes that read on a phone, contrast and text limits, each with its source and the date it was checked, 18 October 2026. One covers the FTC rules on claims, former prices, free offers, urgency, reviews and endorsements, likeness and AI-made people, plus each platform's ad policies. One holds ten angle types and thirteen static formats with layouts, and how to size and read a test. The last is a worked example for a fictional deodorant brand. Claude never invents a claim, price, offer or review; it leaves a marked gap and asks.
Claude does not generate photographs in claude.ai, and it cannot see your ad account or predict results. It writes prompts for your own image tool and reads the results you upload. With code execution on, it counts every character, checks colour contrast, draws wireframes at phone size and builds the upload sheet; without it, counts must be checked in your ad manager. The rules file covers ads shown in the US and flags what a lawyer should look at rather than giving legal advice. You get the most from it by attaching product photos that show every word on the pack, test reports, real customer comments and your past ads with their results.
Claude skill · version 1.0
AI ad creative
Checking your access.
In the download · 8 files
- SKILL.mdThe method Claude follows: brief and proof bank, angles, concepts, image prompts, words, variants, then the check, one stage at a time.
- README.mdHow to install it in claude.ai or Claude Code, how to start, and the licence.
- references/worked-example.mdA full first ad set for a fictional refillable deodorant brand: brief, proof bank, scored angles, three concepts, an image prompt and assembly spec, counted copy for four platforms, a test plan and the check.
- references/angles-and-formats.mdThe proof bank, audience temperature and awareness, ten angle types, angle scoring, the two-channel rule, thirteen static formats with layouts in pixels, carousels, and how to size, name and read a test.
- references/policy-and-claims.mdFTC rules on claims, former prices, free offers, urgency, reviews and endorsements, likeness and AI-made people, FDA cosmetic and drug claims, and Meta, TikTok, Pinterest and Google ad policies, sourced and checked on 18 October 2026.
- references/platform-specs.mdMaster shapes, safe zones in pixels, type sizes that read on a phone, contrast, and text limits for Meta, TikTok, Pinterest and Google, with sources and the date checked.
Install it and begin
01Install it in Claude CodeTerminal mkdir -p ~/.claude/skillsunzip ~/Downloads/ai-ad-creative.zip -d ~/.claude/skills/# Start it with /ai-ad-creative. If this is your first skill, type /reload-skills once.On claude.ai there is nothing to unzip: upload the zip as it is. Every step is on Install and use a Claude skill.
02Your first message to ClaudeText Use the ai ad creative skill. I sell [PRODUCT] to [WHO BUYS IT] in the US. The ads will run on [PLATFORMS AND PLACEMENTS] for a [COLD, WARM OR HOT] audience. The offer is [EXACT OFFER, OR NO OFFER]. Here is what I can prove: [PASTE FACTS, REVIEWS AND SOURCES]. Start with the brief and the proof bank.Attach product photos that show every word on the pack, plus any test reports, price list and real customer reviews. Past ads with their results help most.
Common mistakes
What goes wrong with AI ad creative, and what the skill does instead
Instead ofAsking for forty headlines and picking favourites.
Do thisChoosing three angles that play different motivations, each tied to a fact you can prove, and testing those.
Instead ofLetting the image tool draw the label, the price and the headline.
Do thisGenerating the scene without words, compositing the real product photo and setting every word in the layout tool.
Instead ofCropping one square image into every placement.
Do thisComposing a master for each shape, with words inside the Stories and Reels safe zones: 14% clear at the top, 35% at the bottom and 6% at each side.
Instead ofSetting a 40 px headline on a 1080 px ad because it looks fine on a laptop.
Do thisChecking at phone width: about 66 px or more at 1080 px wide, and roughly twice that on a Pinterest pin, which shows in a two-column grid.
Instead ofWriting "plastic-free" or "was $30" because it sounds right.
Do thisUsing only facts from the proof bank in their allowed wording, and former prices that really were the regular price.
Instead ofA generated person holding the product above a line like "my new favourite".
Do thisA real customer photo with consent, or a disclosure on the image. TikTok needs its AI label, and New York requires synthetic performers in ads to be disclosed.
Instead ofChanging the angle, image and headline at once, then calling a winner after a day.
Do thisChanging one thing per comparison, sizing the test to about 50 optimisation events a week, and reading it on an agreed date.
What to do it with
- ClaudeFree tier, paid above it
A language model that writes, edits, reads long documents and runs tools, including agents that work inside a codebase.
Bad at It is agreeable by default, so it will improve a bad idea instead of telling you it is bad. Ask it to argue against you and the quality of the answer changes.
- Claude CodeSubscription
An agent that reads a whole repository, edits files, runs the tests and explains what it changed.
Bad at It is fast enough to make a large mistake quickly. Work on a branch, read the diff, and never let it push. Watch the starting mode: on Pro, Max and Team plans it opens in auto mode, which Anthropic's own list says may read a .env file and send the key in it to the API that key belongs to. Switch it off.
- PhotoshopSubscription
Still the place a picture is finished. Generative fill, expand and remove now sit on the same layer stack as everything else.
Bad at Generative fill matches its surroundings rather than obeying a brief, so it is a patching tool and not a directing one.
- FigmaFree tier, paid above it
The shared file where interface design happens: components, variables, prototypes and the handoff to code.
Bad at Its generative features make plausible screens rather than considered ones, and a plausible screen is the hardest kind to argue with in a review.
- CanvaFree tier, paid above it
Templates, brand kits and generation in one place, built so a non-designer can produce something acceptable.
Bad at Acceptable is the ceiling as well as the floor. Its templates are so widely used that a trained eye dates a deck in a second.
Why it matters
Asked for ad ideas, a chat returns forty headlines, a few invented statistics and a picture with a garbled label. None of that survives review or teaches you anything once the ads run. This skill makes Claude start from what you can prove, choose three angles that differ in motivation, compose every shape for its placement, keep your product exact and check each ad against platform rules and the FTC before any money is spent.
- Type
- technique
- Period
- ongoing