Creative Ad Analysis

---
name: creative-ad-analysis
description: Analyse ad creatives against performance data to spot what makes ads work. Use this skill whenever a user wants to map visual creative elements (hooks, angles, formats, copy, imagery) to ad metrics like CTR, CPA, ROAS, CPC, CPM, or thumb-stop rate. Trigger for D2C brands, performance marketing teams, or anyone uploading ad images/videos alongside a metrics sheet and asking "what's working", "why is this ad winning", "what patterns do you see", or "help me read my creative performance". Also use when the user wants to brief new creative based on past winners. Do NOT trigger for media buying strategy, audience targeting, or budget allocation questions that have nothing to do with the creative itself.
---

# Creative Ad Analysis

A skill for reading ad creatives next to their performance numbers and pulling out useful patterns — what hooks land, what visual angles convert, what dies on the feed.

This skill exists because the work is harder than it looks. A model can describe a single ad just fine, but the moment you hand it a folder of creatives and a CSV from Meta or TikTok, things fall apart. Naming conventions clash across platforms. Metrics mean different things in different ad accounts. The model starts guessing to fill gaps and the output drifts from useful to wrong.

The point of this skill is to keep the work grounded: clean data first, clear thinking second, pattern-finding third.

## When to use this skill

Use it when the user has both of these:

1. Ad creatives — images, video stills, or video files
2. Performance data — a sheet, export, or pasted table with metrics tied to those creatives

Without both, stop and ask for the missing half. Do not guess metrics from looking at an ad. Do not invent ad descriptions to match metrics.

## The core workflow

Work through these steps in order. Skipping ahead is what causes the drift the post warns about.

### Step 1: Audit the inputs before doing anything else

Before any analysis, check:

- *Are creatives matched to rows?* Each creative needs a clear ID that ties to a row in the performance data. If the file is Ad_v3_final_FINAL.jpg and the sheet says Creative 12, you have no link. Ask the user to confirm the mapping.
- *Are platforms mixed?* Meta, TikTok, YouTube, and Google all report metrics differently. A "view" on TikTok is not a "view" on YouTube. If the data spans platforms, split it by platform before analysing.
- *Are naming conventions consistent?* If half the ads are named by hook and half by date, flag it. Ask the user to standardise or tell you the convention.
- *Are the metrics comparable?* CTR from a brand campaign and CTR from a retargeting campaign are not the same signal. Spend levels matter too — an ad with 50 impressions is noise.

If any of these checks fail, surface the problem to the user before proceeding. Do not paper over messy data. The post is clear: most of the time loss came from the model trying to help when the data was not ready.

### Step 2: Describe each creative on its own

For each creative, tag these elements. Keep tags short and consistent so they group well later.

- *Hook*: what grabs attention in the first 1–3 seconds (for video) or the first glance (for static). Examples: "founder talking to camera", "product close-up", "before/after split", "bold text claim".
- *Angle*: the persuasion approach. Examples: "social proof", "problem/solution", "price anchor", "curiosity gap", "authority/expert", "lifestyle aspiration".
- *Format*: static image, carousel, UGC video, studio video, animation, meme.
- *Copy style*: long caption, short punchy, question, list, testimonial quote.
- *Visual style*: clean studio, lo-fi phone, text-heavy, lifestyle, packshot.
- *CTA*: shop now, learn more, swipe up, none visible.

Keep this tagging plain and literal. Do not interpret intent. Do not guess what the brand was trying to do — just describe what is on screen.

### Step 3: Join tags to metrics

Now line the tags up against the performance data. Look for patterns across:

- Which hooks have the best CTR
- Which angles drive the lowest CPA
- Which formats hold attention longest (thumb-stop rate, video hold rate)
- Which copy styles convert at lower CPC

Only call out a pattern when there are enough data points to back it. Three winning ads with the same hook is a signal worth mentioning. One winning ad is an anecdote. Say so.

### Step 4: Write the read-out

Give the user a short, honest read. Structure it like this:

*What's working*: 2–4 patterns with evidence. Each pattern names the element, the metric it moved, and how many ads back it up.

*What's not working*: 1–3 patterns of failure, same evidence rule.

*What to test next*: 2–3 concrete creative ideas that follow from the patterns. Not vague advice — actual hook + angle + format combinations to brief.

*What I'm not sure about*: things where the data is thin, the platforms got mixed, or the sample is too small. Be plain about uncertainty. The post's main lesson is that pretending to know is worse than admitting the gap.

## Things to avoid

- *Do not invent metrics.* If a number is missing for an ad, say so. Do not estimate.
- *Do not describe creatives the user has not shown you.* If they sent 8 ads and the sheet has 12 rows, only analyse the 8.
- *Do not mix platforms in one pattern.* "Video ads convert better" is meaningless if half the videos are TikTok and half are YouTube. Split first.
- *Do not over-claim from small samples.* Anything under ~5 ads per pattern is a hunch, not a finding. Word it that way.
- *Do not skip the data audit* even if the user is in a hurry. The whole point of this skill is that the audit is what saves time downstream.

## When the data really is too messy

If after the audit the data cannot support analysis — wrong granularity, no creative-to-metric mapping, mixed platforms with no platform column — tell the user plainly. Suggest what they would need to fix before trying again:

- A single export per platform
- A consistent ad name or ID column shared between the creative files and the metrics
- A spend or impression filter so you are only looking at ads with a real test behind them
- One row per creative, not one row per day per creative (or aggregate first)

It is better to send the user back to clean the data than to produce a confident-sounding analysis built on guesses.

## Example output shape

*What's working*

- Founder-to-camera hooks pulled the highest CTR (avg 2.1% across 6 ads vs account avg 1.3%). Strongest on Meta, weaker on TikTok.
- Price-anchor angles ("normally $80, today $39") drove CPA 30% below account average across 4 ads.

*What's not working*

- Studio packshot statics underperformed on every metric across 5 ads. CTR roughly half the account average.

*What to test next*

- Founder-to-camera hook + price-anchor angle in a 15-second UGC video for Meta.
- Lo-fi phone-shot before/after for TikTok specifically — current TikTok ads all use studio footage and none are winning.

*What I'm not sure about*