Creative Diversity Audit

---
name: creative-diversity-audit
description: Audit an ad batch for genuine creative diversity — are your 15-20 ads actually different arguments and formats, or one ad in twenty coats? Use when asked to audit creative diversity, check if ads are too similar, diagnose rising CPMs or narrowing delivery, review a batch before launch, fight creative fatigue at the account level, or someone says 'we have lots of ads but performance keeps dropping'. For generating the missing variety, see ten-angles-copy and hook-writer. For batch production, see bulk-ad-variations.
metadata:
  version: 1.0.0
---

# Creative diversity audit

The algorithm rewards range: give the platform many creatives and it pushes budget
to winners for you — and nobody can pick the winner in advance, so ten real shots
beat one. That was the corpus's argument years ago; in 2026 it is literally how
delivery works. Meta's system reads the creative to choose the audience, clusters
visually similar ads together, and **penalises lazy variations** — a colour swap
and a reworded line are one ad in its eyes, and in this audit's too.

This skill measures whether a batch actually has the diversity everyone assumes it
has.

## Before you start

| Input | Why |
| --- | --- |
| The batch — creatives with their angle/persona labels | The audit walks every asset |
| The naming convention state | persona_angle_variant labels make the audit mechanical; missing labels are themselves a finding |
| Performance data, if live | Symptoms confirm what the matrix predicts |
| The target count | 15–20 genuinely distinct actives is the working band |

## Step 1 — Classify every creative on four axes

| Axis | Distinct means | Lazy variation means |
| --- | --- | --- |
| **Angle** | Different argument (the 10-angle system) | Same argument, synonyms |
| **Hook** | Different opening pain/claim | Reworded same hook |
| **Format** | Static vs video vs carousel vs UGC-style | Same format, new border |
| **Visual skeleton** | Different layout/read-order | Colour swap on one skeleton |

Two creatives are duplicates for this audit when they match on **three or more
axes**. The colour-swap-plus-word-change "variant" matches on all four.

## Step 2 — Build the diversity matrix

Count distinct values per axis across the batch, then the duplicate clusters:

- 20 ads using 2 angles = 2 arguments with 18 echoes — the delivery system sees
  roughly 2 ads.
- The corpus floor: the shortlist tests **15–20 distinct angles**, and versions
  carry **different hooks** — that is the bar for "different".

## Step 3 — Read the symptoms against the matrix (live batches)

| Symptom | What the matrix will usually show |
| --- | --- |
| CPMs climbing week over week | Audience tuning out near-identical ads — fatigue at batch level |
| Delivery concentrating on 2–3 ads | The rest cluster as duplicates of the winners |
| New "variants" die on arrival | They cluster with the fatigued original and inherit its history |
| Strong start, cliff at 2–3 weeks | Normal fatigue — but no genuinely fresh rotation banked |

## Step 4 — Prescribe by gap, not by volume

The fix is never "make more ads" — it is "make the missing kinds":

1. Missing angles → **ten-angles-copy** for the unused arguments.
2. Missing hooks → **hook-writer** bank across the angles present.
3. Missing formats → the format skills (static, carousel, UGC video).
4. One skeleton everywhere → **reference-deconstruction** on a *different* winner,
   or a deliberately new layout.
5. Retire the duplicate cluster's losers; keep the cluster's best performer.

## Output

```
# Diversity audit: <batch> · <n> creatives

## Matrix
Angles used:    <n>/10 — <which>     Hooks distinct: <n>
Formats:        <static n · video n · carousel n · UGC n>
Skeletons:      <n distinct>

## Duplicate clusters (≥3 axes matched)
| Cluster | Members | Keep | Retire |

## Verdict
Effective distinct creatives: <n>  (target band 15–20)
<one line: e.g. "20 assets, 4 effective ads — the batch is an echo chamber">

## Gap prescription
| Missing | Count needed | Route to |
| angles: Enemy, Confession... | 5 | ten-angles-copy |
| video format | 4 | ugc-style-video-script |
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| Every asset audits as unique | Axes applied loosely — synonyms counted as new hooks | A hook is distinct only if the pain/claim differs |
| Nothing audits as unique | Axes applied absurdly strictly | Same angle in static AND video is legitimate diversity |
| Can't classify without guessing | Naming convention absent | That IS a finding: label first (persona_angle_variant), audit second |
| Team disputes "duplicate" calls | Judgement by taste, not axes | Three-of-four matched axes is the line; show the table |
| Batch passes, performance still weak | Diversity was not the problem | Check the offer (offer-builder) and creative quality (meta-ads-creative) |
| Audit passes at launch, fails in a month | Diversity decays as losers get paused | Re-audit at every rotation; the 2–3 week refresh applies |

Never respond to a failed audit by generating twenty more variants of the
survivors. That refills the batch with echoes — the prescription table names what
kinds are missing.

## Rules

- **Three-of-four axes matched = duplicate**, because a bright line beats taste
  in every batch argument.
- **Count effective creatives, not files**, because twenty files with two
  arguments is two ads at delivery.
- **Prescribe by gap**, because volume without variety is what caused the audit
  to fail.
- **Unlabelled batches get labelled before audited**, because traceability is what
  makes diversity measurable at all.
- **Keep one per duplicate cluster**, because retiring everything wastes the
  cluster's learning — the best member carries it.

## Related skills

- **ten-angles-copy** / **hook-writer** — generate the missing arguments and hooks.
- **bulk-ad-variations** — rebuild batches with diversity as a deliberate axis.
- **reference-deconstruction** — a second skeleton when one dominates.
- **meta-ads-creative** — per-creative quality once batch-level variety is fixed.