Building a Cinematic AI‑Avatar Promo — end to end, for any client

1 · How to use this document

You are an AI agent with tool access (a Higgsfield connector or equivalent generation API, a shell with ffmpeg/ffprobe/curl, and the ability to view images). Your job: produce one finished promo video for [CLIENT].

• Work in phases, in order. Each phase produces named files in one working folder. Do not skip QA gates.
• Verify with your eyes and ears, not with assumptions. After every image or video generation, download the file and actually look at a frame. After every audio step, measure it (ffprobe, volumedetect, a spectrogram). Most of our failures were invisible until we looked.
• Never invent a number, ID, price, or fact. If you need the client’s logo URL, their real dashboard numbers, or the founder’s voice sample, obtain the real thing (Phase 1). Placeholder data is fine only when explicitly labelled as illustrative.
• One working folder, e.g. ./promo/ with subfolders vo/ broll/ leo/ logo/ work/ out/. Absolute paths everywhere.
• Preflight every paid generation with the cost tool (it is free) before spending. Watch the credit balance; a low balance can trigger a surprise auto‑charge (see Budget).

2 · The Golden Rules

3 · The Mistake Log

Every one of these cost us time on the Etari build. The rules above exist because of them. Read them so you don’t repeat them.

4 · Pipeline overview & the 4‑layer method

You are producing a ~50–60 second promo: a branded mascot (in the founder’s real cloned voice) delivers a tight 6‑beat pitch, and you cut fast (every ~2–3s) between the talking mascot and supporting visuals. You do not generate 2–3s clips. You generate a few longer pieces and cut between them in the edit, with the VO running continuously underneath. Four layers stack on the voice:

The 6‑beat script skeleton (proven structure): 1) HOOK — a sharp reframe of the problem · 2) PROBLEM — the pain, concretely · 3) SOLUTION — how [CLIENT] is different · 4) PROOF — the dashboard / results · 5) ALT‑OFFER — a secondary path (e.g. DIY) · 6) CTA — the ask + where to go.

5 · Tool stack & model IDs

6 · Phase 1 — Discovery & the script

Gather (real data only)

• Brand: the client’s website URL, the real logo asset (Phase 7), brand colours (sample from the site), fonts.
• Spokesperson: the founder’s name and a clean voice sample (30–90s of them talking — a WhatsApp voice note or a podcast clip works). This is what you clone.
• Substance: the value proposition, the differentiator, one or two real proof points (results, a dashboard screenshot, case numbers). If you can’t get real numbers, use clearly‑illustrative ones and label the dashboard “demo data”.
• Mascot direction: is there a brand mascot/character? (Etari’s logo is a lion inside an apple → “Leo the lion”.) If none, propose one derived from the logo.

Write the 6‑beat script

Tight, spoken, ~50–55s total. Short sentences. One idea per beat. Write it as 6 blocks; each block becomes one VO generation and one BASE mascot clip. Example (Etari):

1 HOOK     "Most businesses don't have an ad problem. They have an agency problem."
2 PROBLEM  "You pay an agency, your account lands on a junior's desk, and once a month you get a pretty PDF."
3 SOLUTION "At Etari Digitals, we do it differently. Every account is run by a founder. No juniors, no handoffs. Google Ads and SEO built as one system, so your ad spend stops being a cost and starts being an asset."
4 PROOF    "And you're never left guessing. Every client gets a live dashboard: spend, revenue, ROAS, profit, all in one place, in real time. No black box."
5 ALT      "Prefer to run it yourself? Our DIY package hands you the exact systems we use internally. Same playbook, your hands on the wheel."
6 CTA      "Want to know what's leaking in your account? We'll audit it free, within 48 hours. Check out the website, link in our bio."

7 · Phase 2 — Mascot & reusable Element

Design the mascot once, lock it forever

Write a canonical design sentence and reuse it verbatim in every prompt. Example: “Anthropomorphic muscular lion, golden‑tan fur, teal‑and‑green mane with a green leaf tuft on top; wears a dark navy athletic polo with thin gold trim and NO text; flat 2D cartoon, bold black line art, cel‑shaded.”

1. Generate the mascot head‑on keyframe with nano_banana_2 using the canonical sentence.
2. Create an Element from that image (a workspace “reference character”). You now have an element_id.
3. In every later prompt, embed the mascot as <<<element_id>>> and still restate the design sentence — belt and suspenders keep it on‑model.
4. Generate the mascot’s set/environment keyframe (the “studio”). For Etari: a dark broadcast studio — charcoal acoustic panels, two vertical white LED strips, a glossy white desk. Save this frame; it is your consistency anchor.

8 · Phase 3 — Clone the founder’s voice

1. Upload the founder’s clean voice sample.
2. Create a voice element (clone) from it. You get a voice_id (element).
3. Do a 5–10 word test generation and listen. If it’s muddy, re‑clone from a cleaner/denoised sample.

The exact VO call (do not deviate)

generate_audio(
  model      = "text2speech_v2",
  variant    = "seed_speech",          # NOT elevenlabs, NOT seed_audio, NOT vibe_voice
  voice_type = "element",
  voice_id   = "[FOUNDER_VOICE_ELEMENT_ID]",
  prompt     = "<the beat's text>"
)

9 · Phase 4 — Generate the VO

• Generate one VO clip per beat (6 clips) with the exact call above. Keep the returned job IDs and download every mp3.
• These raw clips are your clean source of truth for audio. Keep them — you will lay the mastered version of these under the picture, not the audio that Wan spits out.
• Note each clip’s duration (ffprobe). The beat’s lip‑synced video will be built to that length.

10 · Phase 5 — Talking mascot (lip‑sync)

For each beat: animate the mascot keyframe to that beat’s VO with Wan 2.7. The mouth is driven by the audio’s phonemes → real lip‑sync.

The Wan call

generate_video(
  model      = "wan2_7",
  resolution = "720p",
  duration   = <seconds of that beat's VO>,     # 2–15
  aspect_ratio = "16:9",
  medias = [
     { value: "[MASCOT_KEYFRAME_ID]",  role: "start_image" },
     { value: "[BEAT_VO_MEDIA_ID]",    role: "audio_references" }
  ],
  declined_preset_id = "[the id from the preset notice, if any]",
  prompt = "The character speaks to camera, natural mouth movements lip-synced to the audio, subtle head/shoulder motion, staying in the same set. Keep the exact character, framing, lighting. Do not distort."
)

ffmpeg -i beat3_wan.mp4 -lavfi showspectrumpic=s=1000x400:legend=1 spec_wan.png
ffmpeg -i vo/raw_beat3.mp3 -lavfi showspectrumpic=s=1000x400:legend=1 spec_raw.png
# open both; the Wan one has a hazy 12-17kHz band the raw one doesn't.

You now have 6 BASE clips (mascot talking, video‑only usable). Optionally generate 1–2 alternate angles (a tighter dead‑on close‑up, a true side profile) using the “same set, change the angle” trick — these give you free cutaways/jump‑cuts in the edit.

11 · Phase 6 — B‑roll

A) Results dashboard (the PROOF beat)

Generate a photorealistic, minimal dashboard with only a few big, legible KPIs. Fewer elements = zero gibberish. Blur/redact any client name.

generate_image(
  model="nano_banana_2", aspect_ratio="16:9", count=2,
  prompt="Photorealistic clean MINIMAL marketing results dashboard on a monitor in a bright modern office.
          Only THREE big bold legible KPI cards: 'ROAS 6.4x', 'Revenue EUR 41,900', 'Leads +228%',
          and one simple upward green line chart. No tiny cluttered text anywhere.
          The account name at the top is hidden behind a solid grey privacy blur bar. Crisp, credible."
)

Then animate the keyframe (subtle push‑in) with Kling — the on‑screen numbers must stay static/legible:

generate_video(model="kling3_0", duration=5, aspect_ratio="16:9",
  declined_preset_id="[preset id if notice fires]",
  medias=[{value:"[DASH_KEYFRAME_ID]", role:"start_image"}],
  prompt="Subtle slow cinematic push-in on the monitor. Dashboard numbers/chart stay perfectly static and legible. Gentle camera drift, soft background office movement.")

B) The client’s website (real screenshot + FFmpeg push‑in)

1. Open the site in a headless browser; screenshot the relevant section at device scale (or zoom the element up first for a crisp capture). Save a tall PNG.
2. Animate it with a pure‑FFmpeg push‑in (Ken Burns). Text stays pixel‑perfect:

# slow 5s push-in on a cropped 16:9 region of the screenshot, 720p
ffmpeg -loop 1 -i site_packages.png -t 5 -r 30 \
  -vf "crop=1920:1080:0:400,scale=3840:2160,\
zoompan=z='min(zoom+0.0007,1.11)':d=150:x='iw/2-(iw/zoom/2)':y='ih/2-(ih/zoom/2)':s=1280x720:fps=30,\
format=yuv420p" \
  -c:v libx264 -crf 18 -pix_fmt yuv420p broll/web_packages.mp4

C) Scene b‑roll (people/work)

Grounded, believable scenes (someone reviewing dashboards, a team at a desk). Nano Banana keyframe → Kling motion, sound:off. Keep it bright and real — not dark/surreal.

12 · Phase 7 — Logo animation

1. Get the REAL logo (R5). Open the site in a headless browser and read the DOM for the true asset URL — usually something like /assets/logo-mark.png — and download it. Also capture the wordmark: it’s often live HTML text, so zoom the header element up (e.g. transform:scale(6)) and screenshot at device scale for a razor‑sharp wordmark.
2. Upscale the mark to 4K (upscale_image, provide source w/h).
3. Compose a seamless 16:9 frame. Sample the exact background pixel and pad with that hex (R‑“seam”). Center the lockup:

# sample the exact bg colour from a corner, then pad with it (no seam)
read R G B < <(ffmpeg -i logo_4k.png -vf "crop=8:8:2:2,scale=1:1,format=rgb24" -f rawvideo - 2>/dev/null | od -An -tu1 | awk '{print $1,$2,$3}')
HEX=$(printf "0x%02X%02X%02X" "$R" "$G" "$B")
ffmpeg -i logo_4k.png -vf "crop=W:H:X:Y,scale=1560:-1,pad=1920:1080:(ow-iw)/2:(oh-ih)/2:color=$HEX,format=rgb24" logo/logo_frame.png

4. Animate with Kling — insist the logo stays crisp and un‑warped:

generate_video(model="kling3_0", duration=5, aspect_ratio="16:9",
  declined_preset_id="[preset id — the word 'dark' triggers 'IN THE DARK']",
  medias=[{value:"[LOGO_FRAME_ID]", role:"start_image"}],
  prompt="Elegant premium logo sting. The logo (emblem + wordmark) stays perfectly sharp, centered, undistorted.
          Subtle only: a soft golden light-glint sweeps left to right, a gentle glow pulse, faint drifting particles, a very slow zoom-in.
          The logo and text must NOT warp, morph, or change shape.")

13 · Phase 8 — Master the audio

Build ONE continuous VO track from the 6 raw beat clips, then master it. This mastered track is the audio bed for the whole film.

# 1) concat the 6 raw VO clips in order (they are the clean source)
printf "file '%s'\n" vo/raw_beat1.wav vo/raw_beat2.wav ... > vo/list.txt
ffmpeg -f concat -safe 0 -i vo/list.txt -c:a pcm_s24le -ar 48000 -ac 2 vo/bed_raw.wav

# 2) master it — the proven chain. ALWAYS end with -ar 48000 (R4).
ffmpeg -i vo/bed_raw.wav \
  -af "highpass=f=75,equalizer=f=2800:width_type=q:width=1:g=2,\
equalizer=f=10000:width_type=q:width=1:g=4,deesser,\
acompressor=threshold=-18dB:ratio=2:attack=5:release=50,\
loudnorm=I=-16:TP=-1.5:LRA=11" \
  -ar 48000 -ac 2 -c:a pcm_s24le vo/bed_master.wav

ffmpeg -i dirty.wav -af "highpass=f=75,lowpass=f=12000,afftdn=nr=12:nf=-30,deesser" -ar 48000 -ac 2 -c:a pcm_s24le clean.wav

14 · Phase 9 — THE EDIT, in FFmpeg

This is the part DaVinci made miserable. In FFmpeg it is a deterministic script. The model: V1 = the 6 BASE mascot clips concatenated end‑to‑end (video‑only); b‑roll overlays on top at chosen times with crossfades; the mastered VO is the audio. Cut ≈ every 2.5s by choosing when b‑roll covers the mascot.

Step 1 — normalise every clip to 1280×720 / 30fps, video‑only

norm(){ ffmpeg -y -i "$1" -an -r 30 -vf "scale=1280:720:force_original_aspect_ratio=increase,crop=1280:720,format=yuv420p" -c:v libx264 -crf 18 "$2"; }
norm leo/01_beat1.mp4 work/b1.mp4   # ... repeat for all beats and all b-roll

Step 2 — build the BASE spine (concat the beats)

printf "file '%s'\n" work/b1.mp4 work/b2.mp4 work/b3.mp4 work/b4.mp4 work/b5.mp4 work/b6.mp4 > work/base.txt
ffmpeg -f concat -safe 0 -i work/base.txt -c copy work/base.mp4   # this is the mascot, continuous

Step 3 — overlay b‑roll with smooth crossfades

Decide a cut map: for each b‑roll clip, a start second and a duration on the timeline. Overlay each on top of BASE with a short alpha fade in/out so cuts dissolve. Pattern for one overlay (repeat/chain per clip, or script it):

# overlay web_packages from t=22s for 4s, with 8-frame dissolves
ffmpeg -i work/base.mp4 -i work/web_packages.mp4 -filter_complex "\
 [1:v]trim=0:4,setpts=PTS-STARTPTS,fade=t=in:st=0:d=0.27:alpha=1,fade=t=out:st=3.73:d=0.27:alpha=1[ov];\
 [0:v][ov]overlay=enable='between(t,22,26)':x=0:y=0[v]" \
 -map "[v]" -c:v libx264 -crf 18 -pix_fmt yuv420p work/base_1.mp4
# chain the next overlay onto work/base_1.mp4, and so on.

For many overlays, generate this chain programmatically from a cut‑map array (each entry: file, t_start, dur). Keep the VO’s beat boundaries in mind so b‑roll lands on the right words.

Step 4 — lay the mastered VO + optional music, and mux

# video = the fully overlaid picture; audio = mastered VO (+ ducked music bed if you have one)
ffmpeg -i work/picture.mp4 -i vo/bed_master.wav -i music.mp3 -filter_complex "\
 [2:a]volume=0.12,afade=t=out:st=52:d=3[mus];\
 [1:a][mus]amix=inputs=2:duration=first:dropout_transition=0[a]" \
 -map 0:v -map "[a]" -c:v copy -c:a aac -b:a 256k -ar 48000 -shortest out/promo_v1.mp4

15 · Phase 10 — Captions, QA & delivery

Burned captions (optional, cheap, high‑impact)

You already have the exact script per beat and each beat’s start time. Emit an .srt from those, then burn it:

ffmpeg -i out/promo_v1.mp4 -vf "subtitles=captions.srt:force_style='Fontname=Arial,Fontsize=22,Bold=1,PrimaryColour=&Hffffff&,Outline=2,Shadow=0,MarginV=48'" -c:a copy out/promo_captioned.mp4

For word‑by‑word / animated captions, render them in Remotion and overlay the transparent caption layer with FFmpeg.

Final QA gate (do all of these)

• Watch it start to finish. Lip‑sync holds? Cuts land on the beat? No gibberish, no warped logo, no visible client name?
• ffprobe out/promo_captioned.mp4 → confirm 1280×720, 30fps, AAC 48 kHz, ~50–60s.
• ffmpeg -i out/... -af volumedetect -f null - → mean ≈ −16 LUFS‑ish, peak ≤ −1 dB (no clipping).
• Deliver H.264 MP4. Never ship an uncompressed/exotic codec (e.g. “YUV 422 10‑bit” in MP4 won’t open on a normal machine — a real trap we hit).

# safe delivery encode
ffmpeg -i out/promo_captioned.mp4 -c:v libx264 -profile:v high -pix_fmt yuv420p -crf 19 -c:a aac -b:a 256k -ar 48000 -movflags +faststart out/FINAL_[CLIENT]_promo.mp4

16 · Credit budget & cost

On the plan we used, credits cost about $0.05 each (a 1,000‑credit top‑up was $49; auto‑refill runs 18 credits/$1). The entire Etari build — including all the failed experiments while learning — came to ~860 credits ≈ $43. A clean first‑go run following this guide should be well under half that.

17 · Troubleshooting quick‑reference