---
name: ad-creative-signal-analysis
description: Analyse why paid-social ad creatives (Meta, TikTok, YouTube) are scaling or dying, by reading the performance signals the algorithm actually cares about — not how polished the visual is. Use whenever a performance marketer, media buyer, or D2C founder hands over ad creatives with performance data and asks why one is winning, why another is failing, or what to make next. Trigger this skill even when the user has not asked for a deep analysis; if the request is "why is this scaling" or "what should we make more of" and there are creatives plus metrics on the table, route through this skill. Do not let Claude answer from the thumbnail alone — a clean, smiling, well-lit ad can be a dog, and a shaky phone video can be the winner. The algorithm ranks signals, not shringaar.
---
# Ad creative signal analysis
## The problem this skill solves
A media buyer uploads a hundred winning ads and asks an AI why they work. The AI looks at the pictures and says: bright colours, smiling faces, make more of these. The buyer shoots fifty more smiling-face ads, ships them to Meta, and watches CPMs climb while hold rate tanks. The AI described the creative. It did not analyse why the algorithm chose to scale it.
Algorithms like Meta Andromeda do not reward pretty. They reward signals — hold rate, thumb-stop rate, three-second view rate, CTR, engagement depth, watch-through. A phone-shot video with a messy kitchen in the background can beat a studio ad because people stop scrolling to watch it. The studio ad is a rishta photo. The phone video is the actual person.
## When to use this skill
Use it whenever all three of these are true:
- The user has given you one or more ad creatives (images, video stills, video files, or descriptions of creatives).
- The user has given you, or can give you, performance data — even rough numbers. Spend, CPM, CTR, hold rate, thumb-stop, ROAS, hook rate, any of these.
- The user is asking a scaling or diagnostic question. "Why is this working?", "Why did this die?", "What should we make more of?", "Which of these should I kill?"
If any one is true, run this skill. If the user has creatives but no metrics, ask for metrics before analysing. Do not guess from the visual alone. Guessing from the visual is the mistake this skill exists to prevent.
Do not use this skill for pure design feedback ("does this look good?"), for brand review ("is this on-brand?"), or for copywriting tone questions. Those are aesthetic calls and a different job.
## The core method
### Step 1. Get the metrics first, the creative second
Before you look at a single frame, ask for or locate the performance numbers. For each creative you need, at minimum, two of these: spend, CPM, CTR, hold rate or thumb-stop rate, three-second view rate, ROAS. More is better. If the user only sent creatives, say so and ask for the data. Do not proceed on looks alone.
### Step 2. Rank by algorithmic signal, not by looks
Sort the creatives by the signal the platform uses to decide what to scale. On Meta that is usually hold rate and thumb-stop for video, CTR and conversion rate for static. On TikTok it is completion rate and share rate. On YouTube it is watch time and retention curve.
The winner is the one the algorithm likes. Not the one that looks clean.
### Step 3. Compare winners to losers on signals, then look at the creative
Now pull out the top three by signal and the bottom three by signal. Only now do you open the files and look. Ask: what did the winners do in the first two seconds that the losers did not? What pattern of cuts, captions, pacing, hook, or framing separates them? Is the winner loud, weird, raw, abrupt, pattern-breaking? Is the loser smooth, centred, branded, polite?
You will often find the winners are the ones that look worse. That is the point. Note it plainly.
### Step 4. Name the signal pattern, not the visual pattern
When you write up the finding, describe what the creative is doing to the viewer, not what the creative looks like.
Weak finding: "Winners use bright colours and smiling faces."
Strong finding: "Winners open with a jarring cut or a spoken hook in the first 0.8 seconds and hold 42% of viewers to three seconds. Losers open with a logo or a slow pan and lose 70% before two seconds."
The first sentence tells the user to hire a better photographer. The second tells them what to shoot next Tuesday.
### Step 5. Make one falsifiable recommendation
End with one concrete test the user can run this week. "Shoot three phone-style hooks in a messy room, spoken straight to camera, first line under two seconds. Run against your current studio control. If hold rate lifts above X, scale." A recommendation that can fail is a recommendation worth acting on.
### Step 6. Say what you could not see
Be blunt about the limits. If you could not watch the video frame-by-frame, say so. If the user did not share hold rate, say the analysis is weaker without it. If the sample is small, say so. The user is spending real money on your answer. They deserve to know where the answer is thin.
## What a good answer looks like
**User prompt:** "Here are my top 10 and bottom 10 ads from last month with spend and metrics. Why is the top set scaling?"
**Bad answer:** "Your top ads use bright colours, clear product shots, and smiling models. Make more like these."
**Good answer:**
