The Creative Engine — Operating Manual
The runnable SOP for the recursive creative learning loop. One page. Genome in, compounding out. Tag everything, define the winner before you launch, learn across generations, climb the autonomy ladder only as you earn it.
Creative is not a series of ads — it's a learning system: tag every creative at birth (Genome), let Meta's inner loop pick winners, then run your outer loop — dissect winners into elements, bank them in the Creative Memory, and brief the next batch from everything you've learned — so each batch beats the last and out-produces fatigue.
1The Creative Genome — tag every creative at birth
The 9-axis schema. You cannot dissect what you didn't label. Every asset is tagged on these axes the moment it's made — the tags live in the ad name (strict convention) and in the creative tracker (Sheet/Airtable). This is the literal foundation of the loop: without it, the data is lost before you ever read it.
| Axis | Example values |
|---|---|
| 1 Concept / Big Idea (For What) | the 5-minute fix · the honest alternative · join the movement |
| 2 Persona / Audience (For Whom) | budget-conscious parent · time-poor founder · festival-goer |
| 3 Message Angle | pain-agitate-solve · before/after · social proof · status · FOMO · founder story · myth-bust · comparison |
| 4 Hook type (first 3s / frame) | pattern-interrupt · question · bold claim · relatable scene · stat-shock · native/UGC open |
| 5 Format | Video / Static / Carousel / Reel · 9:16 · 4:5 · 1:1 · sound on/off |
| 6 Visual treatment | Hi-fi (polished / studio) vs Lo-fi (UGC / native / raw) |
| 7 Production source | AI-generated · UGC creator · studio · stock · hybrid |
| 8 CTA | Shop Now · Learn More · Sign Up · Get Offer |
| 9 Funnel / awareness stage | cold · warm · hot |
Plus IDs — batch no. variant ID launch date — so every score traces back to an exact element combination.
Sample ad-name string (the genome, encoded):
Read left to right: batch · concept · persona · angle · hook · format · treatment · source · CTA · stage · variant · date. Keep field order identical across every ad so the tracker pivots cleanly.
2The Fitness Function — define "winner" before any batch launches
A winner is decided by the deepest real-money event — never by a scroll metric. Set the thresholds and the data bar before you launch, or you'll crown noise.
① The verdict — deepest money event. CPA / ROAS / cost-per-lead, optimised for the deepest event that matters (media-buying Day 4). This is the only metric that declares a winner. Example bar: CPA ≤ €18 or ROAS ≥ 2.4×.
② Leading indicators — diagnose, don't decide. Hook/thumbstop rate, hold rate, CTR, CVR are the vitals chain (Day 3). They tell you why a creative won or lost — not whether it won. Crown by thumbstop alone and you breed scroll-stoppers that don't sell.
③ Minimum-data bar — anti-noise. No verdict until the variant clears a significance/volume floor — e.g. ≥ ~50 optimisation events and a minimum spend + run-time per cell (media-buying Day 5, learning phase). Below the bar, the result is superstition.
3The Two Loops — never conflate them
Meta runs the inner loop for free. You build the outer loop on top. Most people only ever describe the inner loop and mistake it for a creative system.
Pick among what exists
- Advantage+ Creative + the auction allocate budget among the creatives you already gave the ad set
- Surfaces winners within one generation
- Fast · automatic · you don't control it
- Makes no new creative
Learn what to make next
- Dissects winners into genome elements
- Banks the insight in the Creative Memory
- Briefs the next batch of new creative
- Slower · human/system-designed · compounding across generations
Produce → Launch → [inner loop picks] → Measure → Dissect → Creative Memory → Brief → Produce…
4The Creative Memory — the compounding mechanism (copy this)
A living, append-only knowledge base. Every batch's dissected learnings flow in, so each new brief stands on all prior batches. Competitors can copy an ad; they cannot copy your accumulated map of what works for whom. Four parts — copy the templates verbatim:
5The Batch Brief generator — EXPLOIT + EXPLORE
Every brief is generated from the Creative Memory — not from a blank page. It always has two halves, and a guardrailed ratio between them.
- Recombine & iterate proven elements (from the Proven Elements Library)
- Lower variance · compounding
- Refresh fatiguing winners flagged in the Fatigue Log
- Milk what works, per persona
- Test Open Hypotheses — new element combos
- High variance · mostly fails · grows the map
- Banks new wins into the library and new losses into the Graveyard
- Keeps the pipeline from going stale
Explore/exploit ratio — a guardrail dial. Default split, bounds never <20% explore · never >50% explore:
6The Autonomy Dial — L0→L3, human in or out
How much of the loop the human runs vs. delegates. Automate the mechanical; keep humans on judgment, taste, brand and ethics. You earn your way up as the Creative Memory and track record mature — you never start at L3.
7The Weekly Operating Cadence — the loop on a rhythm
The same loop, on a weekly clock. Fatigue is what forces the cadence — the loop must out-produce decay.
Pivot the creative tracker by genome tags; the Creative Memory and the next brief read straight off it.
8The "Are you compounding?" scorecard
The proof that the engine is a learning system and not just a treadmill of tests. Track these three across batches — if all three move the right way, batch N is genuinely inheriting batches 1…N-1.
Treating each batch as a blank slate. No Creative Memory → no compounding → you relearn the same lessons forever. This is the #1 reason most "creative testing" never compounds.
Going autonomous before guardrails and data exist. Optimising a flawed fitness function at scale = fast, confident garbage. Earn your way up the dial; never start at L3.
The engine in one breath
- Tag at birth — the 9-axis Genome in the ad name + tracker. You can't dissect what you didn't label.
- Define the winner first — deepest money event decides; leading indicators diagnose; minimum-data bar kills noise.
- Two loops — Meta's inner loop picks among what exists; your outer loop learns what to make next.
- Creative Memory — Proven Elements · Graveyard · Open Hypotheses · Fatigue Log. The real moat.
- Brief from memory — EXPLOIT + EXPLORE, 70/30 within 20–50% bounds.
- Climb the dial — L0→L3, automate the mechanical, keep humans on taste/brand/ethics, behind seven guardrails.
- Run it weekly, forever — fatigue is the clock; the loop must out-produce decay. CPA↓, win-rate↑, time-to-winner↓ = you're compounding.