The exact loop top realtors run — now automatic.

Every hour, our miner scans real-estate YouTube across 12 hot US metros — Phoenix, Tampa, Vegas, Nashville, Charlotte, Raleigh, Atlanta, Denver, Miami, Houston, Austin, Dallas-Fort Worth.

We rank by views-per-day, not raw views. A 6-day-old video at 22 vph is what's actually winning right now — not a 3-year-old monster that already saturated.

We dedupe by format (Moving to / Don't move to / Cost of living / Living in) so the 4 we pick for you span 4 distinct video angles.

We pass each viral source's full transcript to a deep-reasoning prompt. The model identifies the hook style, the value-block sequence, the retention tricks, and the emotional arc — stripped of city-specific facts.

The result is a transferable BLUEPRINT: same structure, ready to drop into a different market.

Blueprints are cached per source video — extract once, reuse for every agent who picks that source.

{
  "title_formula":
    "AVOID Moving to [CITY] —
     [HOOK_REASON]",

  "hook_blueprint": {
    "opening_7s": "shock claim
       about [neighborhood]",
    "promise":   "I'll tell you
       which 3 areas to skip"
  },

  "structure": [
    { "STAKES": "30s",  needs_data: false },
    { "PROOF": "30s",   needs_data: true  },
    { "VB_1": "90s",    needs_data: true  },
    { "VB_2": "90s",    needs_data: true  },
    { "PEAK": "60s",    needs_data: true  },
    { "CTA": "30s",     needs_data: false },
    { "LOOP": "15s",    needs_data: false }
  ],

  "retention_tricks": [
    "10s pattern interrupt",
    "B-roll on every beat",
    "Reveal value at 70%"
  ]
}

We pull the last 8 transcripts from your own channel via YouTube's auto-captions — your actual rhythm, your catchphrases, your vocabulary signature, the way you handle numbers and locations.

These transcripts go into the script prompt VERBATIM. No lossy LLM summarization — the writer reads the real you.

If your channel doesn't have transcripts yet, we run a one-time Whisper transcription on a few uploads (5 min, on us).

We don't generate a 12-minute script in one shot — that's how AI hallucinates fake neighborhoods and made-up tax rates.

Instead, every beat (HOOK → STAKES → PROOF → VALUE_BLOCK_1..N → PEAK → CTA → LOOP) gets its own focused call. Prior beats are passed as context for continuity.

Beats that need local data (prices, schools, taxes) call Azure's web_search tool. Every cited fact is stamped with its source URL. Anything unverifiable gets a [verify:] flag for you.

We pull your channel avatar (or upload your own head-shot). Then we hit Azure GPT-image-2's image-edit endpoint with TWO inputs: a top-performing reference thumbnail + your face.

The model copies the reference's composition, typography, palette, and emotion — but uses YOUR likeness as the subject. 4 variations: face-left + bold headline, arrow + red circle, split-screen comparison, big-number callout.

Never generic AI faces. Always the person who has to film the video.