Ad Creative

Who it is for

The owner still running the same three ads from launch, and the performance marketer who wants twenty distinct angles on Monday without a design queue.

What it reads from the brain

Before every run. Sections are addressed as layer.section.

What it does, step by step

1. Read

   Fetches the product page. Extracts facts, price, images and on page reviews. Cross checks every fact against claims_allowed.

2. Angles

   From the angle ladder (problem, mechanism, proof, offer, identity, comparison, honest urgency) and your customers' objections, produces the batch of distinct concepts. No two share an angle plus visual system pair. Each has a hook under 8 words, body under 20, a cited proof line, a CTA, a visual system, and the objection it answers.

3. Claims ledger

   Every hook and body line is checked against claims_allowed and your category's claim restrictions. Anything failing is rewritten or dropped. The ledger is delivered with the batch.

4. Compose

   Background generated from the visual system. The real product image is placed on it, never a generated product. Text set from a layout template. Exported per format (1:1 and 4:5 by default, 9:16 on request). Cost recorded per image.

5. QA

   A second model pass scores each creative on nine dimensions, with your brand name injected so your own logo is not flagged as third party. Under 7 of 9 is regenerated once, then dropped. The QA sheet is delivered.

6. Copy options

   Per creative: 4 headlines, 4 primary texts, 4 descriptions, so you can mix. Desire and offer framing first. A curiosity hook is never in slot one.

7. Deliver

   Contact sheet, individual files, copy sheet, claims ledger, QA sheet.

8. Upload (Act)

   Creates each ad into a paused ad set in a paused campaign named StoreCrew plus the date. Reads back effective_status and fails the upload if anything is not PAUSED. Then one approval: Uploaded n paused ads to Meta. Nothing is running. Activate at this budget? Activation and budget are separate approval gated tools; this agent requests, never executes without the approval.

Tools it uses

• storefront.pageProduct facts, price, images, on page reviews
• image.generateBackground per visual system, hosted endpoints only
• image.composePlaces the real product image on the background
• meta_ads.campaign.create, adset.create, ad.createAct: paused objects in a campaign named StoreCrew plus the date
• meta_ads.campaign.activateApproval gated. Requested, never executed without your tap
• meta_ads.budget.setApproval gated. Requested, never executed without your tap

The three modes for this agent

Guardrails and approvals

Enforced in code. Items with a filled circle wait for your tap in Telegram.

• Never unpauses. Never sets a budget except through the approval gated tool. Never edits a campaign it did not create.
• Read only on your ad account except creating paused objects in its own campaign.
• 40 creatives a month, checked before generation starts. The summary states usage.
• Spend per run capped at $0.30 of generation cost; generation stops when reached.
• The product image is always the real product image. Generated product renders are rejected.
• No claim in any creative that is not in claims_allowed with a source. The ledger proves it.
• No two creatives share angle plus visual system.
• Under 7 of 9 QA after one regeneration is dropped, not shipped.
• Image models through hosted endpoints only. Video through templates only, never customer uploaded code.

What it writes back to the brain

Dated and sourced. Learnings and performance sections are append only.

Example output

Price