💸 No Camera. No Agency. A $3 Founder Ad.

A brand-agnostic video ad production pipeline with four tracks: motion-graphics reels (HTML to MP4, free to run), UGC founder ads (phone footage to a captioned, loudness-normalized cut), cinematic seed-to-video (AI-generated openers and B-roll), and synthetic-UGC talking heads (a HeyGen AI avatar speaks your script in a cinematic 9:16 shot, no camera or studio). Two anti-AI-slop techniques carry it: the bokeh trick that renders every on-screen UI out of focus so generated footage reads as real, and a terminal-style typing keyword-reveal overlay with synced keystroke SFX for punch. One config-driven ffmpeg assembler, one verify gate. Ships a finished spot for 1 to 3 dollars instead of an agency's 500 to 2,000. The unit economics: near-100 percent on Track A (no API spend); 90 percent plus on Track C after the 1 to 3 dollar generation cost margin.

Weekly Teardown: the tool tested in production, what broke, what shipped, the config to try yourself.

What the old way costs: Agency video production still runs 500 to 2,000 per spot. AI video generators work now, except for one giveaway: legible fake UI on screens reads as generated and kills the ad. The bokeh prompt technique erases it entirely. Whisper word-level timing plus two-pass loudnorm assembly turn raw phone footage into broadcast-clean audio with no editor. Video ad production just dropped from an agency line item to a script you run yourself. Synthetic avatars crossed the line where a founder talking-head ad no longer needs a camera, a studio, or your own face on film: a HeyGen avatar delivers the script in cinematic 9:16 format you ship as-is or layer a typing keyword-reveal onto.

What breaks: AI video generators render legible on-screen UI as visibly fake, which kills ad performance instantly.

What shipped: Brand-agnostic and config-driven (no hardcoded values, works for any product). The anti-slop bokeh technique is the actual moat. Track A motion reels cost nothing to run (Playwright plus ffmpeg, zero API spend). Every render passes a deterministic verify gate (ffprobe dimensions plus duration, frame extraction, LUFS level-match under 1 LU). You stop paying per-ad agency rates and stop shipping AI footage that looks AI. Track D generates a founder-style talking-head ad from an AI avatar, so you ship UGC without ever being on camera. The typing keyword-reveal overlay with synced keystroke SFX adds a recognizable on-screen style in one pass.

The config: See The Claude Code Recipe below the paywall for the full production setup, prompts, and schema.

This kit ships a complete Ad Factory operator system. What's included:

• Track A: render_motion_reel.mjs, a Playwright frame-by-frame renderer that seeks a GSAP timeline deterministically into a 1080x1920 H.264 MP4 (free to run, no API calls)
• Track C: seed_gen.py and image_to_video.py, bokeh-anchored prompt to 9:16 seed still (gpt-image-1) to animated 5 to 10s clip (Runway gen4_turbo or fal seedance)
• Track B: compose_vo_ad.py, the config-driven VO-demo assembler with a Whisper timing spine, word-level .ass captions, two-pass loudnorm, BGM ducking, and a LUFS verify gate
• capture_broll.mjs: Playwright scroll-capture of any product page into 9:16 B-roll (headful mode for Cloudflare-protected pages)
• examples/config.example.json: the brand-agnostic config (brand name, colors, accent words, VO path, segment map) that drives the entire assembler
• Track D: heygen_source.py turns a HeyGen AI avatar into a cinematic 9:16 talking-head clip (no camera or studio), ready as a base clip for the typing overlay or the VO assembler
• Typing keyword-reveal overlay: gen_typing_overlay.py, gen_keystroke_sfx.py, and compose_typing_ad.py type key words letter-by-letter with a blinking cursor and synced mechanical-keyboard SFX, burned over any clip