— Door 04 · For the decision-maker

10–15 tests a month.
Without lowering the bar.

The role's real ask isn't "write good copy" — it's throughput: a high volume of on-brand A/B variants, fast, that make the AI smarter every cycle. That's an engineering problem, and I've built the engine. The honest pitch: AI carries the routine 90% — drafting variants, atomising assets, logging results — so my human 10% (the hook, the brand voice, the call on what's actually a winner) goes where it counts.

CONCEPT MODEL — The dashboard figures are an illustrative capacity sketch, not a performance guarantee or Puffy data. Real velocity would be set with your Content Manager and CRO team against the live testing stack. The point is the method.

01The division of labour

One rule runs the whole system: AI never decides what's on-brand or what won a test. It clears the runway; the human lands the plane.

— Carried by the AI engine

The routine 90%

Mechanical, repeatable, verifiable. Speed without touching taste.

▹Drafting copy variantsFirst-pass headlines, body, CTAs for LPs, email, ads, social — in brand-voice Gems.
▹Atomising pillar assetsOne insight → 80+ test-ready pieces across channels.
▹Logging & tagging resultsCaptures every variant + outcome into the test log automatically.
▹Surfacing winning patternsClusters what's converting so the prompt library compounds.

Guardrail: the AI never publishes, never declares a test winner, never sets brand voice. It proposes; I dispose.

— Owned by the human (me)

The decisive 10%

Irreducible — the part the role is actually paying for.

◆The hook & the angleThe single idea a variant lives or dies on — that's judgement, not generation.
◆Brand voice & tasteMaking AI output sound like Puffy, not like AI. The final edit is human.
◆Calling the winnerReading significance vs. noise, deciding what scales and what dies.
◆Compliance & claimsKeeping pricing/health claims defensible (incl. AU strikethrough & clinical-claim law).

Why it stays human: a fast variant with the wrong claim is worse than no variant. Judgement is the product.

02Insight → A/B variant pipeline

Five stages from a customer insight to a shipped, logged test. The tag shows who owns each — so you can see exactly where my judgement lives.

STAGE 01

Insight

Customer pain, data signal or CRO hypothesis identified with the team.

Human + CRO

STAGE 02

Prompt

Insight + brand voice fed into the right asset Gem.

I author · AI runs

STAGE 03

Generate & atomise

Batch of on-brand variants across channels, logged automatically.

AI engine

STAGE 04

Edit & approve

I cut, sharpen the hook, fix voice, check claims — the 10%.

Human only

STAGE 05

Test & learn

Variants ship to A/B; winners feed the library for next cycle.

I call · AI logs

03Capacity model · the command centre

An honest sketch of what this division of labour buys — illustrative figures, not a promise. The number that matters is the last one.

PUFFY · CRO CONTENT ENGINE · CAPACITY SKETCHIllustrative

A/B tests / month

10–15

The role's target velocity — met by systematised variant production.

Variants / pillar insight

80+

Atomisation across LPs, email, ads, social from one idea.

Human time on judgement

~10%

Hook, voice, winner-call — where my hours actually go.

Brand/claims gate

100%

Every shipped variant human-edited & claim-checked. No exceptions.

Throughput scales with the engine; the quality gate does not move. If pace and the bar ever conflict, the bar wins — and I'll say so.

04How velocity never lowers the bar

SG-1 · GATE

Human-edited, always

AI drafts; a variant is never live until I've sharpened the hook and approved the voice. Speed comes from removing my typing, not my judgement.

SG-2 · CLAIMS

Every claim defensible

Pricing and health claims checked against the rules — including AU strikethrough-pricing law and clinical-claim limits — before anything ships.

SG-3 · BRAND

Voice locked in the Gem

Brand voice lives in the prompt library, not in my memory — so volume doesn't drift off-brand and the next writer inherits it.

SG-4 · DATA

Winners, not vanity

Tests called on conversion significance, not clicks or vibes. Losing variants are killed fast; learnings are kept.

SG-5 · COMPOUNDING

The loop gets smarter

Every winning pattern updates the library — the JD's "compounding advantage," built in by design rather than hoped for.

SG-6 · TRACEABLE

Every test logged

Prompt, variant and result captured — so any number is auditable and the CRO team can rerun the reasoning.

The pitch in one line: industrialise the production, never the judgement.

That's the engine. AI carries the routine so the testing velocity Puffy wants is a system, not a scramble — and the human 10% that makes copy convert gets all of me. You've now seen all four doors: the materials, the teardown, the creative, and the engine behind them.

← Back to the hub Open application materials →

10–15 tests a month.Without lowering the bar.

The routine 90%

The decisive 10%

The pitch in one line: industrialise the production, never the judgement.

10–15 tests a month.
Without lowering the bar.