agentleFS
adswipefileadswipefile/adswipefile-plugin/skills/audit-ad-creative/SKILL.md

audit-ad-creative

Skill0 starsChanged 27 days agoMITView on GitHub

Audit one ad creative, your own or a reference from the AdSwipeFile library, for why it lands or doesn't, judged from a stranger scrolling at speed, and return a short ranked feedback note where every problem ends in a fix. Use for "analyze this ad", "why isn't my ad working", "critique this creative", "what is this ad doing", or before iterating on an ad.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 4 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every later judgment is keyed to these.

    line 38
  2. Distinctive:

    Run them silently; they produce findings, not headings.

    line 53
  3. Distinctive:

    No headers, scores, framework names or retelling of the ad.

    line 96
  4. Very specific:

    A reference: when the user asks, save_ad it and set_ad_note with the keeper and the move to borrow.

    line 107
  5. Distinctive:

    Never state an ad's results.

    line 112
  6. Distinctive:

    Keep reads few: one ad, then a small sample of references.

    line 113

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Audit an ad creative

Needs the AdSwipeFile connector. This skill reads the AdSwipeFile ad library of Meta ads through its MCP server. If no AdSwipeFile tools (search_library, list_ads) are available, say so before anything else and give the step for this host. Claude Code: claude mcp add --transport http adswipefile https://adswipefile.com/api/mcp. claude.ai or ChatGPT: add https://adswipefile.com/api/mcp as a custom connector. Other clients: https://adswipefile.com/docs/mcp?utm_source=agent-skill&utm_medium=referral&utm_campaign=audit-ad-creative. Connecting opens a sign-in to an AdSwipeFile account. Steps that need only the user's own material can go on; never invent ads, brands or links in place of the library.

A diagnosis of the creative itself, from the scroller's seat first and the strategist's desk second. It explains what the creative does; it never explains performance from the creative alone. Words such as anchor, persona, funnel job and mechanic are used as the creative-strategy skill's glossary defines them.

0. Look at the actual ad

  • The user's own ad: the image, video frames, copy or a link they share. When nothing came with the request, ask for it. A video seen as one frame plus a transcript is partial: say so, and judge the first frame and the words, not scenes you didn't see.
  • A reference: get_ad for its breakdown units, the first 3 seconds of speech, why_it_works and run. Its run is a reason it's worth auditing, never a result.
  • Context that sharpens it: how the product is really used (ask when the ad doesn't show it), where it runs (cold feed, retargeting, a creator's account), and customers' words the user gives you. A judgment that leans on missing context names what would settle it.

High-production brand film and animation follow partly different rules: say so up front and apply only what transfers (glance, strategy, claims).

1. Classify the job first

Every later judgment is keyed to these. Most bad audits grade a click ad as if it had to close.

  • Anchor (pain or desire, precisely), persona (life context), awareness stage measured against the category this ad sells. Judge stage by what the ad requires the viewer to already accept; blaming a familiar problem on an unexpected cause is educating, so it reads cold.
  • Funnel job: close, click or sort. A click ad may withhold how it works and owes no in-ad CTA; grade the sharpness of the question it leaves. A sort ad is graded on how fast the viewer can place themselves.
  • Distribution: cold in-feed, a creator's own account, or organic to followers. A slow opening that only works with a known speaker is a dependency to name, not a flaw.
  • The floor no job excuses: contradiction. A persona the footage contradicts, a comparison with no named alternative, a claim that trails off.

2. The checks (internal)

Run them silently; they produce findings, not headings.

  • Glance: half a second, phone, sound off, no brand context. Can a stranger tell what kind of product this is and who it's for? The fix is usually the lightest visual element that makes the category read. Process demos and deliberate teases may reveal the product later if its form shows early.
  • Product read: does the crop let the product read as itself; is it used the way customers use it? Showing a use that never happens is a hard flag.
  • Thumbnail: what survives at small size; the point that only works full size doesn't work.
  • First frame, muted (video): one persistent text line carrying the premise; captions don't count.
  • Hook: spoken line, first visual and on-screen text, read together. Aligned, complementary or fighting? Keep "the copy is good" separate from "the pairing is good". Confusion is not curiosity.
  • Decode: for each line, "'line', meaning X". A line you had to decode, the feed won't.
  • Unstated referents: "the secret", "the problem", "the difference" that never appears.
  • One point: list every reading the copy allows; name the one that lands, if any.
  • Setup to payoff: tension built must resolve into the product.
  • Claims: informative or empty; backed or bare; one customers care about. Cold audiences need concrete, backed claims; warm ones may hear confident shorthand. When a line welds a real descriptor to an empty superlative, keep the descriptor and replace the superlative.
  • Native voice: would a customer say it? Elevated words pass when they do a job.
  • CTA, when the job is close: specific, matched to stage, ideally saying who it's for.
  • Mechanic: name it and check it completes (pick-creative-mechanic; list_ads with mechanic finds ads that complete the same move).
  • Format: right container for the angle, mechanic and stage? Does the visual explain or decorate? Comparisons: rows map one to one, both sides styled the same. "Right idea, wrong vessel" names the better format (get_format for how it is built).
  • Body (video): the product answers something set up early; each beat earns the next. A late payoff after an early problem is tension, not a flaw.

Then cluster findings by the fix they point to. Each distinct fix appears once.

3. The feedback note

The default answer, 60 to 150 words, short end preferred:

  • A plain verdict sentence first.
  • What's working, one line per keeper, folded into its fix where possible ("keep the night-shift angle, just show the shift").
  • The biggest problem, named as the biggest, justified from the scroller's seat, with the viewer's literal questions when confusion is the problem ("Is that number good? Who is this for?"), and a sketched fix ("something like: '...'", or the shot to change).
  • The next distinct problem and fix, if any. Stop there.

No headers, scores, framework names or retelling of the ad; point at its parts ("the third line", "the crop"). A strong ad gets small additions, not rescues. Ask for a "full teardown" to get the checks as sections; a batch gets one note per ad plus a line on patterns across the set.

4. Keep what it found

  • The user's own ad: the fix list stays in the reply. For each fix that needs a model, point to reference ads: the examples from get_format for a format change, or list_ads by visual_format, hook or mechanic, with the ad's awareness stage and sort: "longest_running". Read two or three with get_ads, link each by its url, and say what to borrow.
  • A reference: when the user asks, save_ad it and set_ad_note with the keeper and the move to borrow.

Rules

  • Never state an ad's results. A long run is a reason to study it; write "kept running for 94 days".
  • Keep reads few: one ad, then a small sample of references.

Done when the job was classified before any check, every criticism ends in a fix, and the note has each fix once.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.