Reference Deconstruction

---
name: reference-deconstruction
description: Reverse-engineer a winning ad into a reusable creative spec — layout grid, hook placement, proof, colour system, and the job of every element. Use when asked to analyse a winning ad, deconstruct a reference, work out why an ad works, copy a competitor's ad structure legally, turn one winner into a template, study an ad before scaling it, or someone says 'make more like this one'. For producing the next creative from the spec, see meta-ads-creative. For scaling it to 100+, see bulk-ad-variations.
metadata:
  version: 1.0.0
---

# Reference deconstruction

You give it a winning ad. It reverse-engineers **why** the ad works and turns that
into a reusable creative spec — the proven skeleton, rebuilt hundreds of times with
fresh messaging while the structure that earned the win stays intact.

The discipline: every element gets two questions. *What is it?* and — the one that
matters — **what job is it doing, and why is it there?** A deconstruction that only
lists elements produces a lookalike; naming the jobs produces a template.

## Before you start

| Input | Why |
| --- | --- |
| The reference ad, at full resolution | You cannot read hierarchy from a thumbnail |
| Evidence it actually won — spend, duration, or results | Deconstructing a loser templates a loser |
| Whose ad it is — yours or a competitor's | Decides what may be reused (structure) vs what may not (their brand, claims, images) |

If there is no evidence the reference performed, say so and downgrade the output
from "proven skeleton" to "hypothesis to test". Ad Library shows an ad ran, not
that it worked — long-running ads are the stronger signal.

## Step 1 — The six-point teardown

Work through all six, in order:

1. **Layout grid and visual hierarchy** — what the eye hits first, second, third.
   Blur-squint the ad; the read order that survives is the hierarchy.
2. **Hook and headline** — placement and size. What fraction of the canvas does the
   first message own?
3. **Subhead, CTA button, offer badge** — positions, and what each one answers.
4. **Social proof** — ratings, reviews, logos: which type, where, how prominent.
5. **Colour system** — background treatment, product placement, where contrast
   concentrates attention.
6. **The job each element does, and why it is there** — one line per element. If no
   job can be named, note it as probably decorative — and probably droppable.

## Step 2 — Separate structure from skin

| Reusable (structure) | Not reusable (skin) |
| --- | --- |
| Grid, read order, element positions | Their logo, brand colours, fonts |
| The *type* of hook (pain-first, number-first) | Their actual claim or statistic |
| Proof *placement and type* | Their reviews, ratings, celebrity |
| Contrast and attention pattern | Their product photos |

For a competitor's ad, everything in the right column is off-limits — the spec
copies the skeleton, never the skin. Your own winning ad has no such limit.

## Step 3 — Write the spec

The written spec is the deliverable — store it as `deconstruction-<ref>.md` so the
whole batch pipeline reuses it. Fresh messaging swaps in per the matrix (persona ×
angle × offer); the skeleton holds still.

## Step 4 — State what the reference does NOT prove

A winner proves its own combination worked for its audience at its time. It does
not prove the hook caused the win, or that the structure survives a different
offer. Note the strongest untested assumption — that is the first thing to vary in
testing.

## Output

```
# Deconstruction: <reference name/source>
Evidence it won: <spend/duration/results — or "UNPROVEN, treat as hypothesis">

## Skeleton (reusable)
Read order:   1. <element> → 2. <element> → 3. <element>
Grid:         <e.g. top 40% hook, middle product, bottom CTA strip>
Hook type:    <pain-first / number-first / question / ...>
Proof slot:   <type + position>
CTA/badge:    <positions>
Attention:    <where contrast concentrates, background treatment>

## Element jobs
| Element | Job it does | Why it is there |
[one row per element; "no job found" = flag as droppable]

## Skin (do NOT reuse if competitor's)
<their brand, claims, imagery — listed so nobody copies them by accident>

## Untested assumptions
<the strongest thing this win does not prove — vary it first>
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| Recreations look like the reference but perform worse | Skin copied, jobs missed | Redo step 1.6 — name the job of every element |
| Spec produces near-identical ads | Skeleton defined too tightly | The spec fixes structure; messaging must vary per the matrix |
| Reference from Ad Library flops when templated | Running ≠ winning | Demand evidence; long-running is the minimum signal |
| Legal/brand complaint | Competitor skin reused | Only the right column of step 2 was ever off-limits — audit the batch |
| Every teardown finds the same six things | Checklist filled, not thought through | The why column is the work; "it's there for contrast" beats "colour: red" |
| Deconstructed ad works on feed, dies in Stories | Placement-specific structure | Note the source placement; re-spec per format — see format-resizer-spec |

Never template an unproven ad silently. "Make more like this one" assumes "this
one" earned it — check, and say so when it did not.

## Rules

- **Name the job of every element**, because the jobs are the template; the
  elements are just furniture.
- **Structure is reusable, skin is not**, because a competitor's brand, claims and
  images are theirs — the skeleton is what you may learn.
- **Evidence before templating**, because deconstructing a loser mass-produces a
  loser.
- **One spec, many messages** — the skeleton holds still while persona, angle and
  offer swap, or the whole point of deconstruction is lost.
- **State the untested assumption**, because a win proves its combination, not its
  parts.

## Related skills

- **meta-ads-creative** — builds the next creative from this spec (its step 1 is
  the quick version of this skill).
- **bulk-ad-variations** — runs the spec across the full persona × angle × offer
  matrix.
- **creative-diversity-audit** — checks the resulting batch is genuinely diverse,
  not one skeleton in twenty coats.