メニュー
Designs a disciplined ad creative testing system with hooks, isolated variables, and statistical rigor. Use when setting up a testing pipeline or when creative performance is declining.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Structures recurring manager-report 1:1s with a rolling shared agenda that balances check-in, blockers, two-way feedback, and career growth instead of status updates. Use when someone asks "what should I cover in my 1:1s", "build me a 1:1 agenda template", "my 1:1s have turned into status meetings", or is setting up 1:1s with new reports. Do NOT use for general meeting agendas - use meeting-agenda instead; for writing a formal review document, use performance-review-writer; for drafting a specific piece of feedback, use feedback-writer.
Analyzes an A/B experiment to a defensible verdict - sample-size and minimum-detectable-effect math, two-proportion significance testing with confidence intervals, and checks for the traps that fake a win (peeking, sample ratio mismatch, multiple comparisons). Use when someone asks "is this test significant", "did variant B win", "how long should I run this experiment", "what sample size do I need", or shows experiment results and asks whether to ship. Do NOT use for logging and organizing ML training runs - use experiment-tracking instead; for causal questions without a randomized experiment use causal-inference; for general funnel or metric investigation use funnel-analysis or product-analytics.
Drafts a margin-aware 3-email cart-recovery sequence that reminds, handles objections, then incentivizes only as a last touch. Use when a checkout is started but not completed and you need cart-recovery (abandoned-cart) emails, when wiring an abandoned-cart flow in Shopify, Klaviyo, or similar, or when asked to recover lost carts without overspending on discounts.
Structures and drafts argumentative academic essays - a contestable thesis, claim-evidence-analysis body paragraphs, a fairly-engaged counterargument, and a synthesizing conclusion - marking every spot that needs a real citation instead of inventing sources. Use when a student or writer asks "help me outline my essay", "is my thesis strong enough", "structure my argument for this paper", or needs a draft refined for logic and formal register. Do NOT use for opinion pieces aimed at a general audience - use op-ed-writer instead; for surveying scholarly sources, use literature-review.
Runs a full WCAG 2.2 accessibility audit of an interface - keyboard, screen reader, contrast and zoom, forms, and media, in a fixed pass order - and delivers a severity-triaged findings table with concrete fixes. Use when someone asks "is this page accessible", "run an a11y audit on this UI", "will this pass WCAG AA", "why can't a screen reader use this form", or is preparing for an accessibility review, procurement questionnaire, or VPAT. Do NOT use for color-and-contrast-only questions such as palette checks or colorblind-safe data visualization - use color-accessibility instead.
Rewrites a section or page into concise, self-contained answer blocks that AI answer engines and search snippets can lift and cite verbatim. Use when restructuring a blog post, FAQ, docs page, or landing copy to win a featured snippet or get quoted by AI Overviews, ChatGPT, or Perplexity, or when a draft buries its answer in setup. Do NOT use for keyword research, full-article drafting, or technical/crawlability SEO - use a dedicated SEO or copywriting skill instead.
| name | Ad Creative Testing |
| description | Designs a disciplined ad creative testing system with hooks, isolated variables, and statistical rigor. Use when setting up a testing pipeline or when creative performance is declining. |
Most ad creative testing produces noise, not learning. Discipline means changing one variable at a time, running tests long enough, and acting only on statistically meaningful results.
Each test must isolate exactly one variable: hook (first 3 seconds of video or headline), visual format (static vs. video vs. carousel), offer framing (discount vs. free trial vs. guarantee), or call to action. Testing multiple variables at once makes the result uninterpretable. Label every creative with the variable it is testing and the hypothesis it is checking before launch.
The hook is the highest-leverage variable in most paid channels. For video, the 3-second view rate and thumb-stop ratio reveal hook strength before the full watch occurs. For static and search, the headline carries the same weight. Run hook tests before any other variable - a great offer with a weak hook will never be seen. A hook passes if 3-second view rate exceeds 30% on Meta or thumb-stop exceeds 25%; a headline passes if CTR exceeds channel benchmark.
Do not evaluate results until reaching the pre-set sample size. For conversion-rate tests, aim for at least 100 conversions per variant - fewer than that and the result is noise. For CTR or view-rate tests, 2,000 impressions per variant is the minimum. Use a significance calculator before declaring a winner; 95% confidence is the threshold. If budget prevents reaching significance in under 21 days, the test is not worth running as a controlled experiment - use directional data only and label it clearly.
Keep a running log of every test: hypothesis, variable tested, variants, launch date, sample size at close, result, and the decision made. The log is the asset. Pattern recognition across 20+ tests reveals which creative elements consistently move the needle for a specific audience. Without a log, teams repeat failed experiments and lose institutional knowledge.
Run one to three simultaneous tests per channel. Close losing variants as soon as significance is reached. Roll the winner into evergreen rotation and immediately start the next test. Creative fatigue on Meta typically sets in at a frequency of 3-4 - launch new hooks before fatigue hits, not after performance drops. Budget 15-20% of channel spend toward testing at all times.
Testing why a Meta prospecting campaign's cost per acquisition is climbing:
Bad test: Launch three new ads that each change the hook, the offer ("20% off" vs. "free trial"), and the format (video vs. static) simultaneously. One wins on CPA after 4 days with 31 conversions total. Team declares the free-trial video the "winning concept" and rebuilds the account around it. Nothing was learned: the win could be the hook, the offer, the format, or noise - 31 conversions across three variants is far below the 100-per-variant floor, and the platform's auto-optimization had already skewed delivery to the early leader.
Good test: Hypothesis: "A pain-point hook will out-perform our current benefit hook on thumb-stop." Hold the offer, format, body, and CTA constant; produce two variants differing only in the first 3 seconds. Pre-register the decision rule: 2,000+ impressions per variant, evaluate 3-second view rate at 95% confidence. Variant B hits a 34% 3-second view rate vs. 26% for A at significance. Log the result ("pain-point hooks win for this audience"), roll B into evergreen rotation, and start the next single-variable test on offer framing - using B's hook as the new constant.
Produce a creative testing system containing:
Do not let the algorithm decide the winner too early - platform auto-optimization will route spend to an early leader before significance. Do not test concepts you cannot scale if they win. Do not test minor color variations before testing message or offer - the highest-variance variables come first. Do not restart or edit a live test variant mid-flight - edits reset learning and invalidate the comparison; kill it and relaunch as a new test. Do not compare variants launched on different dates as if they ran head-to-head - seasonality and auction pressure confound the result.
