Video ads.
No camera.
HeyGen + Claude Code

Read <client-website-url>: homepage, about, services/pricing. Read what is
actually there, not what you assume. Then write two files.

design-theme.json — pull the REAL hex + fonts off the site, never eyeball them:
{
  "brand":  "#0B0C10",   // dark background + caption-box colour
  "accent": "#4D7CFF",   // the ONE hot colour that does all the pointing
  "fonts":  { "display": "Clash Display", "body": "General Sans" }
}

brand-dna.md — under two pages:
## Verified proof
Every claim you can quote from the site, with its source page: years in business,
client counts, results, guarantees, founder names. RULE: only these numbers may
appear in an ad. Not in this table = does not exist.
## Voice
5 rules with examples, a banned-words list, one sentence this brand would never say.
## Audience
The exact person in one sentence, plus their country and language.

Read design-theme.json and brand-dna.md first. Use design-theme brand as the caption-box
colour and accent as the highlight. Match the Voice rules. Use ONLY numbers from the
Verified proof table.

Use the HeyGen MCP to generate a 20 second vertical video ad, 1080x1920, portrait.

MODE: avatar-less b-roll montage. No presenter, nobody speaks to camera, no lip sync.
People may appear in scenes but never talk to camera.

VOICEOVER (speak exactly, word for word):
"Monthly reporting takes your team four hours a client. Ours takes fifteen minutes.
Campaign builds, days to hours. Forty ads from one brief. AI deletes the boring seventy
percent. Your people keep the judgment. Find your constraint."

SCENES:
[0-2s] HOOK: a marketer alone at night surrounded by spreadsheet tabs, face lit by the
  screen, clock reading late. Boxed overlay (top): "4 hours per report."
[2-6s] the same report assembling itself, charts snapping into a branded PDF, clock
  hands spinning backwards. Boxed overlay: "15 minutes."
[6-11s] a grid of finished ad creatives filling a screen from one short typed brief.
  Boxed overlay: "40 ads. 1 brief."
[11-16s] a strategist at a whiteboard with a client, thinking and talking, not typing.
  Boxed overlay: "Your team keeps the judgment."
[16-20s] solid --brand end card. Centered boxed overlay: "Find your constraint"

TEXT TREATMENT (every overlay): render each line INSIDE a solid --brand rounded caption
box with padding and a small --accent underline. Key word --accent, the rest white.

TEXT SAFE AREA: every box sits fully inside the central 80% of frame, 10% margin from
every edge. Never cropped, never touching an edge, never across a face or mouth.

CTA CARD: final scene is a solid --brand end card, CTA centered, max 70% of frame width.
No bright background, no clipart, no cartoon props, no 3D icons.

NUMBERS: never render a figure, counter, chart or percentage as a generated graphic.
Every number appears ONLY inside a boxed overlay, exactly as written above.

CASTING: take the Audience line from brand-dna.md. State ethnicity, age and setting
explicitly. Never inherit a local-market default for an English or global ad.

PACING: cut every 1.5-2.5s. Pattern interrupt in the first second.
MUSIC: energetic but professional, -20dB.

Read design-theme.json and brand-dna.md first. Use design-theme brand as the caption-box
colour and accent as the highlight. Match the Voice rules. Use ONLY numbers from the
Verified proof table.

Use the HeyGen MCP to generate a 25 second vertical video ad, 1080x1920, portrait.

AVATAR: use avatar_id <your-avatar-id>. This exact avatar speaks the hook and the
CTA to camera. Do NOT substitute a stock presenter.

STRUCTURE (avatar bookends, b-roll core):
[0-6s]   avatar to camera: the hook line.
[6-19s]  b-roll cutaways only, no avatar, voiceover continues over the scenes.
[19-25s] back to the avatar to camera: the CTA line.
Cut away from the avatar every 3-4s. Avatar is on screen under 50% of the total.

VOICEOVER (avatar + VO speak this exactly, word for word):
"Most people selling you AI for your agency have never run one. We do.
<one line of proof from the Verified proof table>. Not a course, not a chatbot,
an operating system we install in thirty days. You own it. Book the constraint call."

B-ROLL SCENES (the 6-19s core):
[6-11s]  a clean terminal on a large monitor, agent logs streaming, a file tree filling
  in. Boxed overlay: "Not a course. An install."
[11-15s] a small calm team shipping finished reports and creatives. Boxed overlay:
  "<proof line, e.g. team of 5, 30 clients>."
[15-19s] one short brief typed, finished deliverables fanning out. Boxed overlay:
  "One brief in. Work out."

TEXT TREATMENT (every overlay): render each line INSIDE a solid --brand rounded caption
box with padding and a small --accent underline. Key word --accent, the rest white.

TEXT SAFE AREA: every box fully inside the central 80% of frame, 10% margin from every
edge. Never cropped, never touching an edge, NEVER across the avatar's face or mouth.

CTA CARD: hold the avatar on the CTA line, boxed CTA caption below the face, never over it.

NUMBERS: never render a figure, counter, chart or percentage as a generated graphic.
Every number appears ONLY inside a boxed overlay, exactly as written.

PACING: b-roll cuts every 1.5-2.5s. MUSIC: energetic but professional, -20dB.

# 1. the END CARD specifically — mid-roll frames lie by omission
ffmpeg -sseof -2 -i final-ad.mp4 -frames:v 1 endcard.png

# 2. a contact strip of the whole ad
ffmpeg -i final-ad.mp4 -vf "fps=1/2,scale=240:-1,tile=6x1" -frames:v 1 strip.png

# 3. the words actually spoken — the check nobody runs, and the one that
#    caught the 45%. Also catches stutters, truncation, invented closers.
ffmpeg -i final-ad.mp4 -vn -ar 16000 vo.mp3
# transcribe vo.mp3, diff against your voiceover

# 4. audio and video should end together — a gap means a cut sentence
ffprobe -v error -select_streams a:0 -show_entries stream=duration -of csv=p=0 final-ad.mp4
ffprobe -v error -select_streams v:0 -show_entries stream=duration -of csv=p=0 final-ad.mp4