Eight stages. One pipeline.
No creative guessing.

Before a single creative gets briefed, our agents run:

• SERP mining for your vertical — what buyers are searching for
• Amazon and Reddit review archaeology — exact language buyers use to describe the problem and results
• Competitor Meta ad library scraping — what's been running 90+ days (the creative that's proven to work)
• Buyer language corpus — the phrases that convert, built from real buyer testimony

Output is a structured research package. Every downstream creative decision traces to something in this package.

The research feeds into:

• Buyer pain matrix — 6–10 core pains ranked by frequency and urgency
• Angle bank — 8–12 creative directions, each tied to a buyer pain and a mechanism
• UMP / UMS — Unique Mechanism Problem / Unique Mechanism Solution — the commercial argument that scales

Jason selects the angle. AI ideates; Jason decides. The approved angle locks the commercial argument for everything downstream.

From the approved angle:

• Copy briefs — one per angle; hook direction, proof points, call to action
• Hook bank — 20–40 tested headline hooks for static and video
• VSL scripts — full video sales letter scripts for video ad production
• Advertorial copy — long-form native content for lander flow

Every piece passes the copy gate before it moves. The gate checks mechanism accuracy, hook-to-research alignment, and call-to-action logic.

• 70+ static image formats per campaign — native-feel, white-background, text-forward, comparison
• Each format briefed with specific visual direction tied to the angle
• QA'd against competitor ad library before shipping — if it's a worse version of what's already winning in market, it doesn't launch

• Talking-head / UGC-style video — AI-generated, briefed from VSL script
• Animated explainer — mechanism-forward; shows how the product works
• Full AI video — Higgsfield, Veo, Seedance, Kling — for concept-forward formats
• VO via ElevenLabs synthetic voice clone

• Full lander build on Cloudflare Workers/R2
• Advertorial-first format — 1,200–2,500 words, editorial style
• Split-test ready architecture — A/B structure is in the build, not added later
• QA'd adversarially against competitor controls before launch
• Message-match QA: does the lander headline mirror the ad hook?

The SkyHouse Meta SOP:

• One ad per ad set — no ad-level ambiguity
• Purchase-event optimization only — no traffic, no engagement campaigns
• 3-stage kill funnel — Day 1, 72-hour, 7-day checkpoints with statistical credible intervals
• Monte Carlo backtests before scaling decisions
• Four in-house buyers — Calvin, Ryan, Matt, Fernando — each owns their portfolio
• Daily buyer reports; escalation to Jason on anything material

• BigQuery + RedTrack + payment processor reconciliation
• Canonical daily ROAS and CPA by offer — not Meta-reported, reconciled
• Early-winner algorithm — flags statistical promise before full spend confirmation
• Kill algorithm — flags underperformers before budget bleeds
• Weekly feedback: what the data says about the angle, the creative, and the funnel — feeds back into Tranche 1