agentleFS
XuHo-ITXuHo-IT/Content-Agent-Kit/skills/ads-creative-lab/SKILL.md

ads-creative-lab

Skill8 starsChanged 4 months agoMITView on GitHub

Design and read Facebook/Instagram creative tests for a content-agent-kit agent — hook × angle × format matrix, a naming convention results can be pivoted on, test structure and budget (ABO test → CBO/Advantage+ scale), kill/scale rules with sample-size floors, and the next round built from the winners. Use when the user says "test creative", "test mẫu quảng cáo", "nên chạy mẫu nào", "làm thêm mẫu ads", "hết ý tưởng creative", "mẫu nào thắng", "scale mẫu nào", or after ads-report/ads-fatigue-watch names a creative to replace. Works from pasted Ads Manager numbers; better with the meta-ads MCP. Never creates or changes ads on its own.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

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

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

    Ask for, or read from KNOWLEDGE.md / mkt-context if installed:

    line 34
  2. Distinctive:

    If none of these exist, say so and start with a 2×2 test, not a plan.

    line 39
  3. Very specific:

    Write plan.json (shape in examples/ads-samples/creative-plan.json) and run:

    line 49
  4. Very specific:

    Every angle needs a proof.

    line 59

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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

Creative lab — test creatives on purpose, not by feel

On Meta in 2026 the algorithm finds the audience; the creative is the targeting. The lever a fashion team actually controls is what the first three seconds say and show. This skill turns "make more ads" into a test that produces an answer.

Hướng dẫn từng bước cho người dùng

Chạy lệnh node scripts/… ở thư mục gốc của kit — thư mục có AGENTS.md, hai cấp trên file này (nếu skill được cài bằng install-skills.mjs --kit, đường dẫn nằm trong file .kit-root cạnh file này).

Dẫn người dùng theo đúng các bước dưới, bằng lời đơn giản. Trước khi bắt đầu: một câu nói skill sẽ làm gì, mất khoảng bao lâu, cần người dùng chuẩn bị gì. Đầu mỗi bước báo "Bước n/N — <tên bước>" kèm một câu đang làm gì. Khi cần người dùng làm gì, chỉ hỏi đúng một việc. Kết thúc: kết quả nằm ở đâu và việc nên làm tiếp.

BướcAgent làm gìNgười dùng cần làm
1. Gom thông tinHỏi sản phẩm (mã, giá, ưu đãi), số liệu đã chạy, ngân sách test và chi phí/đơn chấp nhận đượcGửi danh sách sản phẩm + số liệu cũ (dán hoặc CSV)
2. Lập ma trậnViết plan.json (hook × góc × định dạng) và chạy node scripts/ads/creative-matrix.mjs plan.json --max 12Chọn hook/góc muốn thử, hoặc để agent đề xuất
3. Soát nội dungChạy từng mẫu chữ qua ads-policy-check—
4. Đề xuất cấu trúc testĐưa cách chia nhóm quảng cáo (ABO), ngân sách mỗi nhóm và cách tính ra con số đóĐồng ý hoặc chỉnh ngân sách
5. Đọc kết quả vòng trướcSo số theo tên quảng cáo, chỉ kết luận khi đủ ≥ 2.000 hiển thị và ≥ 3 ngàyDán số liệu sau 3–5 ngày chạy
6. Dừng — người dùng tự lên adsKhông tự tạo, sửa, tắt hay đổi ngân sách quảng cáoTự dán tên quảng cáo vào Ads Manager, đúng nguyên văn

Hay vướng:

  • Đổi tên quảng cáo cho "dễ nhìn" → kết quả không gộp được theo hook/góc nữa; giữ nguyên tên script sinh ra.
  • Mới chạy 1–2 ngày đã đòi biết mẫu nào thắng → agent sẽ ghi "chưa đủ dữ liệu", chờ đủ ngưỡng.
  • Không có số liệu cũ → bắt đầu bằng một test nhỏ 2×2 thay vì kế hoạch lớn.

0. Inputs

Ask for, or read from KNOWLEDGE.md / mkt-context if installed:

  • product(s) with a short code (VL01 = váy linen dáng A), price, the real offer
  • what has already run and its numbers (paste, CSV, or meta-ads MCP)
  • monthly test budget and the cost per result the business can afford (target CPA)

If none of these exist, say so and start with a 2×2 test, not a plan.

1. Build the matrix — three variables, nothing else

variablewhat it isfashion examples
Hook (H)first 0–3s, words + picturecâu hỏi tình huống · POV · "1 món 3 cách phối" · try-on xoay người · so sánh outfit A/B · comment khách làm hook
Angle (A)the one reason to buychất vải/độ mát · dáng che khéo · giá/ưu đãi · đổi size miễn phí · phối được nhiều kiểu · review khách thật
Format (F)containervideo 9:16 try-on · carousel phối đồ · ảnh đơn flat-lay · UGC/KOC · catalog/collection

Write plan.json (shape in examples/ads-samples/creative-plan.json) and run:

node scripts/ads/creative-matrix.mjs plan.json --max 12          # variety-first subset
node scripts/ads/creative-matrix.mjs plan.json --max 12 --csv > matrix.csv

Names come out as AW26_VL01_H1-A2-V9_261111. Paste them into Ads Manager verbatim — the whole point is that next week a pivot on the name splits results by hook, angle and format.

Every angle needs a proof. "Vải mát" with no shot showing it is a claim; run each draft through ads-policy-check before it goes anywhere.

Producing the assets: video cells → create-video (genres flash-deal, review, testimonial), image/carousel → design-campaign, UGC → ugc-koc-brief.

2. Test structure (propose, do not build)

  • Test campaign: ABO (budget per ad set), 1 ad set per hook or per concept, 3–4 ads each, broad audience (age/country only) so the creative does the targeting. Same optimisation event as the scale campaign (purchase if the pixel/CAPI sees ≥ ~50 purchases a week, otherwise add-to-cart or messaging).
  • Budget per ad set: enough for ~50 link clicks or ~3× target CPA within 3–5 days. State the number you computed and the assumption behind it.
  • Scale campaign: CBO or Advantage+ shopping; winners move in as new ads with the same post ID (keeps social proof), not by raising the test budget 5×.

3. Read the test — floors before verdicts

Per ad, only judge after both: ≥ 2,000 impressions and ≥ 3 days (or spend ≥ 2× CPA).

signalreadsmeans
hook rate3-second views ÷ impressions (video)does the opening stop the scroll
hold rateThruPlays ÷ 3s viewsdoes the body keep them
CTR (link)link clicks ÷ impressionsdoes the angle make them want it
CPA / ROAScost per purchasedoes the page + offer close
  • Kill: spend ≥ 2× target CPA with 0 purchases; or CTR in the bottom third of the test after the floor.
  • Keep testing: passes the floor, CPA within 1.3× target.
  • Scale: CPA ≤ target over ≥ 3 purchases. Fewer than 3 purchases is not a winner — write "not enough data", exactly as ads-report does.

Diagnose by layer, not by ad: good hook rate + poor CTR = angle problem; good CTR + poor CPA = landing/offer problem (price, size out of stock, ship fee) — not a creative problem.

4. Iterate from winners

The next round changes one variable of the winner: same angle with 3 new hooks, or same hook in 2 new formats. Record each round in brain/creative-tests.json:

{ "round": 3, "date": "2026-11-11", "winner": "AW26_VL01_H1-A2-V9_261101",
  "learned": "POV hook beats question hook on cold traffic (CTR 2.1% vs 1.2%, 6k vs 5k impressions)",
  "next": ["H5-A2-V9", "H6-A2-V9", "H1-A2-U1"] }

learned names the number and the sample. A lesson with no sample is a rumour.

5. Stop and ask

Output: the matrix, the proposed campaign/ad set structure with budgets, and the read of last round. Then stop. Creating, duplicating, pausing or re-budgeting ads on a live account is the user's call, even with the meta-ads MCP connected — see docs/18-ads-and-marketing.md.

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.