Skip to content

Makes

Ad variants

Writes ad variants per theme inside each platform's limits, then checks them for length, message match, repeated angles and claims without proof.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick ad-variants.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
Responsive search ad limits
Needs
The themes, with pillar, proof, offer and channel · The platforms to write for
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedPaid media
SkillPaid media

Ad variants

Example output

Write search ads for one theme
HeadlineLengthCall
Pipeline Attribution Software29Pin 1
Live in 14 Days on Average26Proof
The Best Attribution Tool25Needs proof

Writes ad variants per theme inside each platform's limits, then checks them for length, message match, repeated angles and claims without proof.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

ad-variants/SKILL.md

In the folder

  • ad-variants/SKILL.md
  • ad-variants/references/platform-limits.md
  • ad-variants/scripts/check_ads.py
---name: ad-variantsdescription: >-  Writes ad variants for each theme inside each platform's own text limits (Google and Microsoft responsive search  ads, LinkedIn, Meta, ChatGPT Ads, Reddit), gives every variant one job and one variable, then checks the set for  length, message match with the landing page, repeated angles, competitor trademarks and claims without proof.  Uses voice-check and claim-check as its check and returns READY, READY WITH GAPS, NEEDS PROOF or BLOCKED. Use it  after ad-themes, or when a new ad group needs ads. Not for choosing what the ads say (use ad-themes) or for judging  the creative and visuals (use ad-creative-critic).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Reads the landing page if the agent can  browse; otherwise the reader pastes its headline. Python 3 is optional, for the length and overlap checks.metadata:  author: SHAKEUP  version: "1.0.0"  stage: make  plugin: paid-media  homepage: https://www.shakeuphq.com/skills/ad-variants--- # Ad variants Writes the ads for each theme so they fit the platform, match the page and test one thing at a time. ## Purpose - One job: turn themes into ad variants per platform, and check them.- Stage: make. It ends with its own check, at most 2 fix rounds.- It never writes a claim that isn't in `## Claims and proof`, never puts a competitor's trademark in ad text, and never uploads or launches an ad. ## Use when / not | Situation | Use || --- | --- || Themes are set and need ads | this skill || A new ad group from `search-terms-triage` needs its ads | this skill || Refreshing ads rated Low for 30 days | this skill, with the old assets as input || Deciding what the ads should say | `ad-themes` first || Reviewing images, video or the whole ad creative | `ad-creative-critic` || Reviewing the landing page itself | `landing-page-critic` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || Themes: name, pillar, proof, offer, channel | yes | `ad-themes` handoff; the reader | ask for one theme's parts (who it's for, the problem, the result, the proof, the call to action); then `BLOCKED` || Platforms to write for | yes | the themes' channels; `## Channels`; the reader | ask once || Landing page per theme (its headline, its offer) | no | browse it; the reader pastes the headline | message match `not checked`; open item || Keyword theme per search ad group (the words the queries share) | no | `search-terms-triage` handoff; the ad group's keywords; the reader | the keyword check is skipped for that ad group || Old assets and their ratings | no | an asset report export | write fresh; nothing is carried over || Voice rules and words to avoid | no | `## Voice` | voice checked on the generic rules only, coverage `partial` || Approved claims, never-say list, competitor names | no | `## Claims and proof`; `### Never say`; `## Competitors` | every number is `unsourced`; competitor check runs on names the reader gives | ## Profile read From `company-profile.md`: `## Company` (brand terms), `## Buyers` and `### Buying group` (who each ad speaks to), `## Messaging house`, `## Voice` (point of view, tone, words to use and avoid, lines that sound right), `## Claims and proof` and `### Never say`, `## Competitors` (names and mention policy), and every `## Learnings` entry for `ad-variants`, `voice-check` and `claim-check` with status `noted`, `fact` or `accepted`. No profile: ask in one message for who the ad is for, the problem, the result, one proof with its source, the call to action and the words the company never uses. Say that `company-profile` saves this next time. ## Steps 1. **Apply learnings.** Asset patterns that won or lost before (a proof headline that beat a feature headline) change the mix in step 4. Name each one that fires. 2. **Load the limits** from `references/platform-limits.md`. If a limit was checked more than 6 months ago, open its source again and say what changed. 3. **Collect the parts per theme:** who it's for (in their words), the moment or problem, the result as a verb phrase ("see pipeline by channel"), one proof with its source, the offer, the landing page headline, the keyword theme for search. 4. **Write per platform.**    | Platform | What to write | Roles each asset plays |   | --- | --- | --- |   | Google or Microsoft responsive search ad | 8 to 12 headlines, 3 to 4 descriptions per ad group (under 8 limits combinations; over 12 and assets never gather impressions) | at least 3 carry the keyword theme (pin 1 of them to position 1, nothing else pinned); at least 2 proof; at least 2 for whom or a qualifier that turns the wrong searcher away; at least 2 the action or offer. Descriptions: the moment and the mechanism; the proof and the action; the objection answered. Every headline must read next to any other |   | LinkedIn single image | 3 variants per theme | the first line of the intro text carries the whole message; the headline names the offer; a call to action from LinkedIn's list |   | Meta feed image | 3 variants per theme | the problem in the first line; proof next; the headline names the offer |   | ChatGPT chat card | 2 or 3 cards per theme, plus context hints | the title follows on from the buyer's question; the body gives the next step, not the answer again; hints describe the use cases and needs, not keywords |   | Reddit | 2 promoted posts per theme | a plain headline a member would write; a useful offer that works without a demo; no ad words | 5. **Vary one thing.** Within a test, each variant changes one variable from the theme's control: the angle, the proof, the offer or the opening. Label each variant with the variable it changes. Search assets are a pool, not variants: their variety comes from the roles in step 4. 6. **Match the page.** Search: headline 1 carries the keyword theme's words, and the headlines share at least half of the landing page headline's words (fewer and the visitor lands on a different promise). Social and ChatGPT: the offer the ad names is the first thing the page delivers, in the same words. 7. **Run the checks.** `python3 scripts/check_ads.py --ads ads.json` (input format in its docstring): lengths against the limits, counts, near duplicates (60% shared words), repeated openers, the keyword theme, message match, competitor names, dynamic keyword insertion in competitor campaigns, dashes, words to avoid, empty words, superlatives and generic calls to action. Without Python, count characters for every asset (double-width characters count 2 on search), compare each pair of assets in a field for shared words, and say you checked by hand. Then run `voice-check` and `claim-check` on the set (inline without them: no em or en dashes, no profile words to avoid, every number and result in `## Claims and proof` and in date, nothing from `### Never say`). ## Check and verdict | Check | Fails when || --- | --- || Length | any field over a hard limit; past a recommendation is a warning || Count | outside the platform's allowed range; outside 8 to 12 headlines or 3 to 4 descriptions is a warning || Variety | in a search asset pool, or in the field a social test varies, 2 assets share 60% or more of their words, or 3 open with the same 2 words (fields a test holds constant as its control are meant to repeat) || Message match | headline 1 lacks the keyword theme; under half the landing headline's words appear || Proof | a number, result, quote or superlative without a verified claim (`claim-check`) || Voice | a line fails `voice-check` || Competitors | a competitor name in ad text; keyword insertion in a competitor campaign || One call to action | an ad asks for 2 things | | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No theme and no message parts after one question || `NEEDS PROOF` | The reader requires a claim that `claim-check` marks unsupported, unsourced, expired or forbidden, or a competitor trademark in ad text: those variants are not written, and the reader is asked for the source or a change || `READY WITH GAPS` | Variants written, but checks still open after 2 rounds, message match not checked (no page), or a platform's limits unconfirmed (each named) || `READY` | Every check passes | Named stops: `BLOCKED` asks for the parts. `NEEDS PROOF` stops on the claim, shows what `claim-check` found, and offers variants built on the approved proof instead. No other stops. ## Loop 1. Run the checks.2. Fix only what failed: shorten, replace a duplicate with an asset that plays a missing role, swap a claim for an approved one, rewrite in the profile's voice.3. Check again. At most 2 rounds, then hand over with what is still open, labelled `open`. ## Learn Append to `## Learnings` when: - an asset report after 30 days shows which assets rated Best or Low: a `fact` naming the role each played (proof, qualifier, action);- a test between variants reads on qualified leads or opportunities (not clicks): a `fact` with the variable that won;- an ad is disapproved: a `fact` with the platform's reason, so later runs avoid it;- the reader rejects a variant's tone: a `proposal` for `voice-check`'s rules. Example: `- 2026-10-24 | ad-variants | fact | Headlines with the 14-day setup proof rated Best in 3 of 4 ad groups; feature headlines rated Low | Why: Google Ads asset report, Sep 2026 | Status: noted` ## Output ```Verdict: <READY | READY WITH GAPS | NEEDS PROOF | BLOCKED> · Coverage: <complete | partial (what's missing)> · Checks: <script | by hand> For each theme and platform:<Theme> · <Platform> · landing page: <headline> · keyword theme: <words>| Field | Text | Length / limit | Role or variable | Pin |Message match: <share or not checked> · Variety: <pass or findings> Not written: <claims or names refused, and why>Check: <each check, pass or fail> · Rounds: <n>Open: <labelled items>Learnings applied: <list, or none>``` Then the handoff block. ## Hand off ```handoffskill: ad-variantsversion: 1.0.0verdict: READYcoverage: completefor: <themes and platforms>outputs:  sets: 3  platforms: [google_rsa, linkedin_single_image]  headlines: 10  descriptions: 4  not_written: []next: [ad-creative-critic, landing-page-critic]open: []``` ## Failure modes | Problem | Do this || --- | --- || The brand name is long and eats the headline | Keep the brand out of most headlines; the display URL already carries it on search || The proof needs a sentence of context | It doesn't fit an ad: pick a proof a buyer can check in one line, or move it to the page || The landing page says something different from the theme | Flag it as the largest fix; message match beats any wording change || The reader wants 15 headlines | Allowed by the platform; warn that past 12, assets rarely gather enough impressions to be rated || Microsoft autogenerated assets are on | Say they're on by default and may write claims you haven't approved; recommend turning them off or reviewing them || A limit can't be confirmed (Reddit) | Write the copy, mark its length `unchecked`, ask for the current limit | ## Influenced by - Each platform's own ad specifications, listed with their dates in `references/platform-limits.md`, for example Google's responsive search ad limits: https://support.google.com/google-ads/answer/7684791?hl=en- SHAKEUP's B2B Google Ads playbook, for the asset roles (keyword theme first, proof or qualifier next, one action), pinning position 1 only, and testing one variable at a time.- SHAKEUP's ad copy checks: fit, who it's for, the result, proof, the next step, you before we, and no words any company could say.

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

Most weak ads are well written but aimed at the wrong stage, or carry a claim nobody can check. So every asset gets one job and one proof from your approved claims, and every set is checked against the page it sends people to.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Read your learnings first

    Asset patterns that rated well or badly before change the mix it writes.

  2. 02

    Load the limits, with their dates

    Every limit comes from the platform's own spec page, dated. A limit older than 6 months is checked again.

  3. 03

    Give every asset a job

    On search, some headlines carry the keyword, some the proof, some the buyer and some the action. Only the keyword headline is pinned.

  4. 04

    Change one thing per variant

    On LinkedIn, Meta and ChatGPT, each variant changes one variable from the control and says which.

  5. 05

    Match the page

    The first headline carries the query's words, and the ad shares at least half the page headline's words.

  6. 06

    Check, fix and check again

    A script counts lengths and overlap, then the voice and claim checks run. Failures get fixed, at most twice.

What you need

  • Required

    The themes, with pillar, proof, offer and channel

    The handoff from ad themes → Your answer

  • Required

    The platforms to write for

    The themes' channels → Your company profile → Your answer

  • Optional

    The landing page headline and offer per theme

    It reads the page if your agent can browse → The headline you paste

  • Optional

    The keyword theme for each search ad group

    The handoff from search terms triage → The ad group's keywords

  • Optional

    Approved claims, voice rules and competitor names

    Your company profile

What it can return

  • READY

    Every variant fits, matches its page and passes the voice and claim checks.

  • READY WITH GAPS

    Variants are written, but the page or a platform's limits couldn't be checked.

  • NEEDS PROOF

    You asked for a claim without a source, or a competitor's trademark in the ad text.

  • BLOCKED

    There's no theme and no message to write from.

Where it stops for you

  • When there's no theme or message to write from
  • When a required claim has no source, or would put a competitor's trademark in the ad

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run for a made-up B2B SaaS company that sells attribution software. The theme is a board asking which channels create pipeline, for one Google Ads ad group.

Verdict

READY

Coverage: complete

Checks: script

FieldTextLengthRolePin
HeadlinePipeline Attribution Software29 / 30Keyword theme1
HeadlineB2B Pipeline Attribution24 / 30Keyword theme
HeadlineBuilt on Your CRM's Own Stages30 / 30Proof
HeadlineLive in 14 Days on Average26 / 30Proof
HeadlineFor Sales-Led B2B SaaS Teams28 / 30Qualifier
HeadlineBook a Demo11 / 30Action
DescriptionSetup takes 14 days on average, and your lifecycle stages stay as they are. Book a demo.88 / 90Proof and action
Message match

the headlines share 5 of the 6 words in the landing page headline.

Not written

"The Best Attribution Tool". It's a superlative with no proof behind it.

The handoff block the next skill reads
skill: ad-variants
version: 1.0.0
verdict: READY
coverage: complete
for: Pipeline attribution ad group
outputs:
  sets: 1
  platforms: [google_rsa]
  headlines: 10
  descriptions: 3
  not_written: [the best attribution tool]
next: [ad-creative-critic, landing-page-critic]
open: []

What good looks like

  • Every asset within its platform's limit, with the length shown
  • The query's words in the first headline, and the page's words in the ad
  • No 2 assets that say the same thing in different words
  • Proof only from your approved claims
  • Each social variant labelled with the one thing it changes

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
    Download ad-variants.zip.
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Which platforms does it cover?

Google and Microsoft responsive search ads, LinkedIn single image, Meta feed, ChatGPT chat cards and Reddit promoted posts. Each limit is dated and linked to the platform's own spec page.

Why does it pin only one headline?

Pinning the keyword headline to the first position keeps the query's words in front of the searcher. More pins cut the combinations the platform can test.

What if Reddit's limits can't be confirmed?

It still writes the posts, marks their length as unchecked and asks you to paste the current limits from Reddit Ads Manager.