agentleFS
itsryanarmaniitsryanarmani/skills/ad-research-to-creative/SKILL.md

ad-research-to-creative

Skill1 starsChanged 9 days agoMITView on GitHub

Plan paid ads for any site or product from public competitor data: pull live Meta ads, rank by run length, pick angles and a persona, write scripts, build landing pages and short video ads, then hand off to a small test. Use before spending on ads for a new project, or when asked to research existing ads, make a spokesperson persona, or build a VSL or landing page per ad.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

14 rules. About 0 are common, 10 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈10

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Check policy risk. A ban on an unverified Meta portfolio stops every site.

    line 14
  2. Distinctive:

    Use 8-10 keywords: the buyer's words, the builder's words, and the education words ("guide", "checklist").

    line 18
  3. Distinctive:

    Look for the nearest guide or lead-magnet ads.

    line 23
  4. Very specific:

    Follow ad-experiments: one audience, one geo, one creative per test.

    line 36
  5. Distinctive:

    Keep outfits polished and fitted.

    line 43
  6. Distinctive:

    No data says that one hair or eye color sells more.

    line 44
  7. Distinctive:

    Use only claims the product contains.

    line 50
  8. Distinctive:

    Do not clone a competitor's people, marks or exact copy.

    line 51
  9. Very specific:

    Use ad-creative-generation for still images and ad-creative-templates for layouts.

    line 52
  10. Distinctive:

    Only true proof: real counts, real excerpts, named public sources.

    line 66
  11. Distinctive:

    Do not promise results. Promise the order of work and the documents.

    line 68
  12. Very specific:

    Verify pixel and Conversions API events: ViewContent, Lead, InitiateCheckout, Purchase.

    line 74
  13. Very specific:

    Read the pixel with GET /{pixel_id}/stats?aggregation=event&start_time=<unix>&end_time=<unix> (give a time range).

    line 75
  14. Distinctive:

    Check business verification before spend.

    line 76

≈ “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

ad-research-to-creative

One repeatable path from "we should run ads" to a paused, tracked test. Each step has an output file. Do not skip to creatives.

0. Check the economics first

  • Compute cold CAC = CPC / click-to-purchase rate. Compare it with gross profit per order, not with revenue: take the price, then remove payment fees, refunds, the cost of making or delivering the product, and any other per-order cost. A digital guide keeps most of its price, but fees and refunds still count.
  • Assume a guide under $100 will not pay back on the first order, and check that with your own numbers. Plan the ladder: free checklist, guide, kit, back end.
  • Mark every benchmark as an assumption until the first $300 of spend replaces it.
  • Check policy risk. A ban on an unverified Meta portfolio stops every site. Start with low-risk niches.

1. Research live ads (read-only, about $0.10)

Script: ad-research OUTDIR US 3 "keyword" ... from the scripts repo (treg, about $0.0019 a page). Run it from a clone of that repo.

  • Use 8-10 keywords: the buyer's words, the builder's words, and the education words ("guide", "checklist").
  • Keyword search is loose. Filter the result to ads that name the topic. Half the raw hits are noise.
  • Meta hides spend on commercial ads. Run length is a weak signal. An ad can stay active while it loses money, so use it to rank ads and pick which to study, never as proof that an ad pays.
  • Group by advertiser. Count ads per advertiser, formats and CTAs.
  • Deep-pull one advertiser's full ad set with Apify apify~facebook-ads-scraper (about $0.005 an ad, free plan has a $5 cap). Use view_all_page_id=<page_id> in the Ad Library URL.
  • Look for the nearest guide or lead-magnet ads. They are the closest model for the offer.
  • The advertisers are also the B2B prospect list: a business that pays for lead ads can pay for a Featured listing.

Auth notes:

  • Keys come from the environment (APIFY_TOKEN, TREG_TOKEN) or a local env file kept outside the repo. Never print them.
  • The treg MCP key can expire while the treg CLI still works. Use treg call, treg balance.
  • A free-plan Apify token still works for small pulls.

Output: summary.md, ads.csv, builder-prospects.csv.

2. Pick angles

  • Name the patterns that run long: pain hook ("Tired of..."), local callout, one-day or one-step promise, free tool, price anchor.
  • Pick 3 to 5 angles. Write one hypothesis each: audience, hook, offer, what result means a win.
  • Follow ad-experiments: one audience, one geo, one creative per test.

3. Persona

  • One spokesperson per site. A guide host, never a fake customer.
  • Image: WaveSpeed wavespeed-ai/jib-mix-qwen-image/text-to-image ($0.02). Prompt: no name tag, no text, alone in frame, no hands. Naming the persona in the prompt makes a name tag appear.
  • Make two options. Review before any video spend.
  • Identity comes from the image, not the prompt. A prompt rebuilt from a chosen image gives the same style but a different face. Once the owner picks a face, keep that image as the reference, in a private place. Make new poses, outfits and backgrounds with the edit model ($0.02, images must be an array). The recipe is in personas/persona-recipe.json.
  • Keep outfits polished and fitted. Meta rejects ads that read as sexually suggestive, and repeated rejections can restrict the ad account. A warm smile and eye contact carry the appeal.
  • No data says that one hair or eye color sells more. Test two faces against each other.
  • Faces look alike across sites unless the prompt changes age, hair and style. Vary them.
  • Video: HeyGen (check the vendor's current API version first) or WaveSpeed lipsync (about $0.10 for a 5 s clip at 480p and about $0.15 at 720p; the model accepts 5 to 20 s of audio per clip and its billing floor can differ from that, so price a clip with wavespeed price and measure one real run from the balance). A personal brand should use the real person, not a character.

4. Scripts and creatives

  • Script shape, 20 to 35 seconds: hook (2 s), pain, one fix, proof from the guide itself, free next step.
  • Use only claims the product contains. No invented testimonials, stats or "worth $X" stacks.
  • Do not clone a competitor's people, marks or exact copy. Copy the pattern, not the words.
  • Use ad-creative-generation for still images and ad-creative-templates for layouts.

5. Landing page per angle

Build the page the ad sends people to, before you build the ad. A bare product page loses the visitor the ad just won. Capture the current page at phone width (430 px) first, and compare.

Two destination types:

  • Free guide (Lead): headline that repeats the ad hook, the video, one email field, the full list of what is inside, a quiet link to the paid offer.
  • Paid offer (Purchase): the sales page below.

Order on a phone: kicker, headline, one sentence, video, button. The headline must be on the first screen, so do not put the video first.

Sales page order: hero (headline, video, button, three ticks) -> trust strip (counts and sources that are true) -> the cost of the problem (four short pains) -> before and after -> what is inside (one card per document, with what it does) -> a real excerpt from the product, as text -> the guide host in two sentences -> options side by side (free, mid, full), with the full one marked -> questions (objections only) -> final button. A sticky button bar on phones.

Rules:

  • Only true proof: real counts, real excerpts, named public sources. No invented testimonials, no fake scarcity or countdowns.
  • The price goes on the page, never in the video.
  • Do not promise results. Promise the order of work and the documents.
  • Match the ad hook in the first line.
  • Flag owner decisions, such as currency.
  • Preview locally at 430 px and at desktop width before anything goes live. Check that the buy form, the free form and every link work in the real build.

6. Tracking, then launch

  • Verify pixel and Conversions API events: ViewContent, Lead, InitiateCheckout, Purchase. The event names, ids and tweaks are in ad-conversion-hub (section "Event ladder for a small ad test"). Key names and credentials are in ad-platform-credentials.
  • Read the pixel with GET /{pixel_id}/stats?aggregation=event&start_time=<unix>&end_time=<unix> (give a time range). Fire a test event first.
  • Check business verification before spend. The status is in Business settings, Security centre.
  • Build the campaign paused. Show the owner the exact settings. The owner approves the budget.
  • Start with a small test, about $300 to $600 for a guide priced under $100. Set the stop line before launch from the economics: the rules differ by goal. A Lead test stops when spend reaches about 3 times the target cost per Lead with zero Leads. A purchase test stops when spend reaches about 3 to 5 times the profit left on one order after all per-order costs with zero purchases, even if some Leads came in, and it also stops if Leads come in but none buy after the same spend. The owner's cap wins when it is lower.

Worked example

A directory site, 2026-11-10: about 800 live ads read for one topic, 33 contractor advertisers found, one free-guide ad in the whole set, and no guide ads aimed at the premium segment. Keep project names, products and hooks in a private notes file, not here.

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.