Skip to content

Checks

Ad creative critic

Checks an ad set for an early hook, early brand, one message, proof and a match to the landing page, then counts the real ideas once near-copies merge.

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-creative-critic.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
ABCD guidelines for YouTube ads, Diversified creative, 3 copy tests
Needs
The ads, with their images or video and their copy
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedCreative
SkillCreative and conversion

Ad creative critic

Example output

Check an ad set
CheckAdsResult
Real concepts3 of 5Pass
The page makes the ad's offer1 adFail
Brand on the creative early4 adsPass

Checks an ad set for an early hook, early brand, one message, proof and a match to the landing page, then counts the real ideas once near-copies merge.

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-creative-critic/SKILL.md

In the folder

  • ad-creative-critic/SKILL.md
  • ad-creative-critic/references/checks.md
  • ad-creative-critic/scripts/concept_clusters.py
---name: ad-creative-criticdescription: >-  Checks a set of B2B SaaS ads (static, carousel, video or text) for an early hook, early brand, one clear message,  proof, message match to the landing page and real variety across the set, where near-duplicate concepts count as  one. Returns pass, fail or not run per check, severities, fixes ranked by impact, a count of real concepts, and a  PASS, PASS WITH NOTES, FAIL or BLOCKED verdict. Use it before ads launch or when a set stops performing. Not for  writing the brief (use creative-brief), for reviewing the landing page itself (use landing-page-critic) or for  search term waste (use search-terms-triage).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Reads ad images, scripts and exports;  reads landing pages if the agent can browse. Python 3 is optional, for counting real concepts.metadata:  author: SHAKEUP  version: "1.0.0"  stage: check  plugin: creative-conversion  homepage: https://www.shakeuphq.com/skills/ad-creative-critic--- # Ad creative critic Checks every ad in a set, then checks the set: how many real ideas it holds once near-copies are merged. ## Purpose - One job: check one ad set (one campaign or ad group) per run, ad by ad and as a whole.- Stage: check. It ends in a verdict with severities.- It never judges creative from metrics alone, never quotes a benchmark the reader didn't give, and never fixes an ad with a claim the profile doesn't support. ## Use when / not | Situation | Use || --- | --- || Ads are ready to launch | this skill || A set's results are falling and you suspect the creative | this skill, with the performance export || Several ads look different but you suspect they are one idea | this skill (S1) || You need a brief for new ads | `creative-brief` || The landing page needs its own review | `landing-page-critic` || Copy variants within platform limits | `ad-variants` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The ads: images, video, and their copy (intro text, headline, button) | yes | a connected ad account; an export of the ads with image files; screenshots or pasted copy | ask once; metrics or ad names alone are `BLOCKED` || For video: the timing | no | the video if the agent can watch it; a script or storyboard with timestamps; a transcript | judge from order; timing `approximate` || The landing page each ad sends to | no | the URL, read in a browser; the first-screen copy pasted; a `creative-brief` handoff | A5 `not run`; coverage `partial`; verdict can't be `PASS` || Platform, placement and objective | no | the export; the reader's note | infer from the format; say so || Labels per ad: angle, hook type, visual concept, offer | no | the `creative-brief` handoff; the reader's sheet | label them yourself in step 2 and show the labels || Performance by ad (impressions, clicks, view rates, frequency) | no | a connected ad account; an export | judge the creative alone; nothing marked `confirmed by data` | ## Profile read From `company-profile.md`: `## Company` (brand terms), `### Buying group` (roles and objections), `## Buying moments`, `## Messaging house` (pillars, to label angles), `## Voice` (for rewrites and humour), `## Claims and proof` and `### Never say`, `## Competitors` (mention policy), `## Channels`, and every `## Learnings` entry for `ad-creative-critic` with status `noted`, `fact` or `accepted`. No profile: ask in one message for the brand name, who the ads are for, and the figures the company can use with their sources. Say that `company-profile` saves this next time. Without approved claims, every first-party figure in the ads is `unsourced`. ## Steps 1. **Apply learnings.** A learning that changes a rule (a hook type that never works for this company, a different variety bar) overrides the defaults. Name it when it fires. 2. **Inventory the set.** One row per ad: id, format, placement, the hook (first line, image headline or first 5 seconds), headline, offer, button, landing page. Label each ad's angle (the pillar or idea), hook type (problem, trigger, outcome, proof, question, contrast, story) and visual concept (comparison, product screen, person to camera, single figure, and so on). Show the labels: S1 depends on them. 3. **Check each ad.** Run A1 to A9 from `references/checks.md`, using its B2B table for how the attention guidelines apply. Quote the ad for every fail. For video without timestamps, judge from the order of the script and mark timing `approximate`. 4. **Check the claims (A4).** Run `claim-check` on the copy of every ad. Without it: list each figure, comparison and quote, and match it to `## Claims and proof`; unmatched is `unsourced`, anything on `### Never say` is `forbidden`. 5. **Check the match (A5).** For each ad, put its promise and offer next to its landing page's first screen. The offer must be the same thing (a tour is not a sales call; a report is not a demo). Ads sharing one page are checked against it once each. 6. **Count real concepts (S1 to S5).** Run:    ```   python3 scripts/concept_clusters.py --csv ads.csv   ```    with columns id, hook, headline, angle, hook_type, visual, format, offer. Without Python: group ads with the same angle, hook type and visual concept; then compare the hook and headline of each pair and join any 2 that differ only by swapped or reordered words. Each group is 1 concept. Required: at least min(3, half the ads rounded up), because 3 is the smallest set where a winner reads as a pattern, and a small set needs half its ads to be different ideas. Say which way you counted. 7. **Rank and fix.** Order fails by severity, then by how many ads share the fail. For every critical and major, write the fix:   - Copy fixes: the rewritten line, in the company's voice, from the messaging house, buyer phrases and approved claims only. Run `voice-check` on the rewrites and `claim-check` on any with a figure; fix and recheck, at most 2 rounds; label what still fails. Without those skills: no dashes, nothing from words to avoid, the 3 copy tests, every figure traced to `## Claims and proof`.   - Variety fixes: name the concepts the set is missing, from pillars and buying moments no ad covers, as briefs for `creative-brief`. Keep the strongest near-copy in each concept and pause the rest (S5), even when S1 passes.   - Match fixes: change the ad's offer or point the ad at a page that makes that offer. 8. **Use the data, if given.** Mark a finding `confirmed by data` when the export agrees (the ad that fails A1 has the lowest view-through or click rate in the set). Never raise a severity from an outside benchmark, and never call an ad good on metrics alone. ## Check and verdict Per ad: `PASS` (no critical or major), `FAIL` (any critical or major). | Set verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No creative to look at (only names, ids or metrics), or video with no file, script, storyboard or transcript || `FAIL` | Any critical finding; or a major on more than half the ads; or S1 fails || `PASS WITH NOTES` | Majors on half the ads or fewer, or minors only; or no fails but A5 was `not run` || `PASS` | No fails and every check ran | Named stop: `BLOCKED` asks for the creative. No other stops: missing labels are filled by the skill and shown; a missing landing page leaves A5 `not run`. ## Learn Append to `## Learnings` in the profile: - Results come back for a concept (cost per lead, demo rate or pipeline by ad from the reader's ad account and CRM): a `fact` naming the concept, the result and the source.- The reader rejects a finding ("our buyers respond to the logo-first open"): a `proposal` to change that check for this company, with the reason.- 2 or more results show a hook type winning or losing for this company: a `proposal` to prefer or avoid it. Example: `- 2026-10-24 | ad-creative-critic | fact | Problem hooks beat question hooks in paid social: 4 opportunities against 0 | Why: the ad account export and CRM deals, Q3 2026 | Status: noted` ## Output ```Verdict: <PASS | PASS WITH NOTES | FAIL | BLOCKED> · Coverage: <complete | partial (what's missing)>Ads: <n> · Real concepts: <n> of <n> required · Formats: <list> · Counting: <script | by hand> · Learnings applied: <list or none> Set| Check | Result | Evidence | Fix |(S1 to S5; S1 lists each concept and the ads in it) Per ad| Ad | Verdict | Fails (id, severity) | Evidence (quoted) | Fix | Top fixes, in order:1. <fix> | ads: <ids> | why first: <severity, ads affected>... Concepts to brief next: <angles or buying moments no ad covers>Rewrites checked: voice-check <verdict>, claim-check <verdict> (<or: inline>)Not run: <checks and why>``` Then the handoff block. ## Hand off ```handoffskill: ad-creative-criticversion: 1.0.0verdict: FAILcoverage: partial (no landing page for 2 ads)for: <campaign or ad set>outputs:  ads: 6  real_concepts: 2  required_concepts: 3  critical: 1  failing_ads: [<id>, <id>]  concepts_to_brief: [<angle or moment>]next: [creative-brief, landing-page-critic]open: [landing pages for <ids>]``` `creative-brief` reads `concepts_to_brief`. When A5 fails because the page is wrong, `landing-page-critic` reviews that page. ## Failure modes | Problem | Do this || --- | --- || Only metrics and ad names | `BLOCKED`: creative can't be judged from its results; ask for the images or copy || Labels from the reader disagree with what the ads show | Use what the ads show; list each relabel so S1 can be checked || One ad in many sizes | 1 ad: check it once, note the sizes under A7 || Dynamic or AI-assembled ads (the platform mixes headlines and images) | Check each asset and the riskiest combination: a headline whose claim needs proof paired with an image without it || A competitor named in an ad | Check `## Competitors` mention policy; outside the policy is `forbidden` under A4 || The set is for a retargeting audience that already knows the brand | A2 can be `minor`; say so; the rest stands | ## Influenced by - Google's ABCD guidelines for YouTube ads (Attract, Brand, Connect, Direct), built on consumer video, used here with the B2B changes in `references/checks.md`: https://www.thinkwithgoogle.com/_qs/documents/8472/ABCD_Complete_V7b_HR_1.pdf- Meta Blueprint's course on diversified creative, which splits variety into concept, motivator and format: https://www.facebookblueprint.com/student/path/253130- Harry Dry's 3 copy tests, used in A1 and the rewrites: https://www.demandcurve.com/newsletters/growth-newsletter-202- What we changed: near-duplicates are counted as one by a rule anyone can check, the set needs a minimum of real concepts, consumer urgency tactics are dropped, and every ad is matched to its landing page's first screen.

How it thinks

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

What we built

Ad sets often look varied and hold one idea. We count near-copies as one by a rule anyone can check, ask for a minimum of real concepts, and match every ad to its page's first screen. Consumer attention rules are kept where they fit B2B and dropped where they don't.

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

    List every ad and label it

    Each ad gets its hook, offer, angle, hook type and visual concept. The labels are shown, because the variety count depends on them.

  2. 02

    Check each ad

    The problem or the result must come in the first line or the first 5 seconds, with the brand on the creative early. Each ad needs one message, and proof for any claim.

  3. 03

    Match each ad to its page

    The page's first screen must make the same promise and the same offer. A tour in the ad and a sales call on the page is a critical fail.

  4. 04

    Count the real ideas

    Ads with the same angle, hook type and visual, or with near-identical wording, count as one concept. Its script does the count, and the skill can count without it.

  5. 05

    Apply consumer rules with care

    Attention rules from consumer video still help, such as an early hook and early brand. Fake urgency is dropped, because B2B buyers compare vendors for months.

  6. 06

    Rank the fixes

    Critical first, then by how many ads share the problem. Rewrites pass the voice and claim checks, and missing concepts become briefs.

What you need

  • Required

    The ads, with their images or video and their copy

    A connected ad account → An export of the ads with the image files → Screenshots or pasted copy

  • Optional

    For video, the timing

    The video, if your agent can watch it → A script or storyboard with timestamps → A transcript

  • Optional

    The landing page each ad sends to

    The live page, if your agent can browse → The first-screen copy, pasted → A creative brief handoff

  • Optional

    Results by ad

    A connected ad account → An export

What it can return

  • FAIL

    Any critical finding, a major on more than half the ads, or too few real concepts.

  • PASS WITH NOTES

    Majors on half the ads or fewer, or minor findings only. Also when the landing page couldn't be checked.

  • PASS

    Nothing failed and every check ran.

  • BLOCKED

    There's no creative to look at, only names or metrics.

Where it stops for you

  • When there's no creative to look at

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 security questionnaire software. The set has 5 paid social ads, all linking to a page that offers a 3-minute tour.

Verdict

FAIL

Coverage: complete

Real concepts: 3 of 3 required

Counting: script

ConceptAdsWhy they count as one
Stop chasing questionnairesA1, A2, A3Same angle, hook and screen. Only a word and the background change
Deals wait on security reviewA4Its own angle and a timeline visual
Answers from your approved libraryA5Its own angle and a product screen
AdVerdictFailsFix
A5FailCritical: the ad offers a demo, the page offers a tourOffer the tour in the ad
A4FailMajor: the logo shows on the last card onlyPut the name on the first card
A1 to A3PassMinor: near-copies split the budgetKeep A1, pause A2 and A3
Concepts to brief next

the security lead's worry that answers go out of date, which no ad covers yet.

The handoff block the next skill reads
skill: ad-creative-critic
version: 1.0.0
verdict: FAIL
coverage: complete
for: same-day questionnaires campaign
outputs:
  ads: 5
  real_concepts: 3
  required_concepts: 3
  critical: 1
  failing_ads: [A5, A4]
  concepts_to_brief: [answers go out of date]
next: [creative-brief, landing-page-critic]
open: []

What good looks like

  • A count of real concepts, with the ads in each
  • Every ad matched to its landing page's offer
  • Every figure in an ad traced to your approved claims
  • Video timing marked approximate when it came from a script
  • Missing concepts named, ready to brief

Install

Use it in the tool you already have.

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

  1. 01
  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.

How does it decide 2 ads are the same idea?

Same angle, same hook type and same visual concept, or hooks and headlines that share most of their wording. A colour change or a swapped word doesn't make a new concept.

Why use guidelines built on consumer ads?

The attention rules hold up, such as showing the problem and the brand early. Urgency tactics don't fit long B2B buying, so it drops those.

Can it read our results?

Yes, as evidence for a finding it already made. It never calls an ad good or bad on metrics alone, and never compares them with outside benchmarks.

What does the script do?

It groups near-duplicate ads and counts the real concepts and formats. It's optional. Without Python, the skill groups the ads the same way and says so.