MakeRoll/Case 04/AI UGC ad campaign/Built 23 Sep 2026
Four AI UGC ads for a hotel-rate site. Every decision attached.
Booking Wolf says every hotel and tour has two prices. One prompt turned four of the price gaps it publishes into photoreal selfie-style ads, and the agents kept the whole record. Prompts, rejected frames, gate scores and a blind review. Members also get the bill.
Members or partners onlytotal metered spend, probe included
85.45 sof finished cuts
3 of 4pass the 100-point gate at 95+
Read this first
A MakeRoll production study of four AI UGC ads for bookingwolf.com. Booking Wolf hasn’t endorsed this page, and none is implied. The ads haven’t been run as paid media, and no audience, click or booking data exists for them. Every on-screen person is an AI-generated fictional character. The scripts adapt example prices and customer-quote copy that bookingwolf.com publishes on its homepage (captured 23 Sep 2026), and MakeRoll hasn’t verified any price. Every number below comes from the project’s ledgers, QA files and review log.
01 / The films
Four stories. One price gap each.
They’re ordered by the blind reviewer’s ranking, best first. Each one dramatises a price example that bookingwolf.com prints in its own marketing. Press play for sound. Captions are on by default.
A · BangkokGate 100.00
Maya is an AI-generated fictional character, not a real traveller, customer or testimonial.
Dramatised example. The prices are example figures from bookingwolf.com’s homepage copy (captured 23 Sep 2026) and are not verified. MakeRoll didn’t check them against Viator or any live booking.
D · RomeGate 99.57
No experiment was run. The twin test is a scripted story.
The twins are AI-generated fictional characters, not real travellers, customers or testimonials.
Dramatised example. The prices are example figures from bookingwolf.com’s homepage copy (captured 23 Sep 2026) and are not verified. MakeRoll didn’t check them against Hotels.com or any live booking.
C · TuscanyGate 95.00
Sofia is an AI-generated fictional character, not a real traveller, customer or testimonial.
The upgrade is a story device, not a promised benefit. The 20 percent commission line and the “saved $340” end card are unverified.
Dramatised example. The prices are example figures from bookingwolf.com’s homepage copy (captured 23 Sep 2026) and are not verified. MakeRoll didn’t check them against Booking.com or any live booking.
B · LisbonInterim
Interim cut. It fails the 100-point gate at 81.86 because two shots were never generated, so Booking.com, €156 and Booking Wolf appear only on the cards.
Clara and Ines are AI-generated fictional characters, not real travellers, customers or testimonials.
Dramatised example. The prices are example figures from bookingwolf.com’s homepage copy (captured 23 Sep 2026) and are not verified. MakeRoll didn’t check them against Booking.com or any live booking.
01of 04
Final cutGate 100.00 / 100 passReview rank 1 of 4
Same Guy
Bangkok · whisper-vlog mystery
On a Bangkok street-food tour, Maya finds out the girl beside her paid $22 less for the same guide. Then she finds out why.
23.78 slength
100.00100-point gate
6.5/10blind review, first cut
Members or partners onlymetered spend
Final cut. Passed the gate at 100.00 with every floor intact.
Blind review, first cut
The reviewer’s pick of the four. A selfie close-up, a clear same-tour, two-prices story, and the only first cut with two concrete prices on screen, $89 and $67 at 3.5 s. It drags in s2 and ends on a weak, generic card.
Identical twins book the identical Rome hotel room for the identical dates. One used Hotels.com, one used Booking Wolf, and the loser sleeps away from the window.
21.23 slength
99.57100-point gate
6/10blind review, first cut
Members or partners onlymetered spend
Final cut. Passed at 99.57 after a mute in the edit removed an unscripted “mm-hmm”.
Blind review, first cut
The premise most likely to stop a scroll, and it says the brand earliest, at 7.9 s. But it opens on both twins blinking, the first cut’s price card had no Booking Wolf number, and the room changes from one bed to two.
Sofia books the cheapest room at a Tuscan farmhouse and gets handed the key to the suite. The script credits the commission the owner didn’t pay.
24.19 slength
95.00100-point gate
5/10blind review, first cut
Members or partners onlymetered spend
Final cut. It passes at exactly 95.00 with the voice check at 0 of 5, so it has zero margin.
Blind review, first cut
A strong key-dangle hook in a pretty world. But the brand doesn’t appear until 16.4 s, the payoff line contradicts itself (she kept it and gave it back), and the first cut’s end card read like a promised suite upgrade.
Interim cutGate 81.86 / 100 failReview rank 4 of 4
Balcony
Lisbon · two-hander neighbour skit
Two strangers on neighbouring Lisbon balconies realise they have the same room and the same view. They didn’t pay the same price.
16.25 slength
81.86100-point gate
3/10blind review, first cut
Members or partners onlymetered spend
What’s missing and why
Interim cut. Shots s3 and s4 were refused with HTTP 402 because the shared treg balance ran too low. So Clara’s price line and the brand line were never generated, the brand is never spoken, and the cut runs 16.25 s against an 18 s minimum. It fails the gate at 81.86. Finishing it needs two more clips.
Blind review, first cut
The most photoreal frames of the set. But frame 0 is a woman in three-quarter profile looking off-screen, the key number is missing from the captions, there are 4 s of silent croissant and coffee, and the brand is never said.
The gate re-measures the delivered file. The reviewer watched frames at 2 fps and read transcripts, without audio, and scored the first cuts before the edit fixes. They agree on the order. They don’t agree on how good the ads are.
Film
Gate /100
Review /10
Hook
Story
Talent
AI tells
Brand
CTA
Length
Spend
A · Same Guy
100.00
6.5
7
7
7
6
7
5
23.78 s
Members or partners only
D · Twin Experiment
99.57
6
6
7
7
6
6
4
21.23 s
Members or partners only
C · The Upgrade
95.00
5
7
5
7
5
4
4
24.19 s
Members or partners only
B · Balcony
81.86
3
4
3
7
7
3
5
16.25 s
Members or partners only
Two of the reviewer’s 23 defects hit all four cuts. The first cuts’ burned-in AI line sat inside the TikTok and Reels overlay zone, and it’s since been removed. All four also share one end-card template with no product shot and a “free sign-in” footer, which is still open.
AI tells is scored so that higher means fewer visible tells. Masters are 1080 × 1920 H.264 at 24 fps with AAC 48 kHz audio at -14 LUFS. The players stream 720 × 1280 web copies of the same cuts, with loudness within 0.3 LU of the masters.
02 / The site and the brief
Two prices for one room. That’s the pitch.
Booking Wolf, Inc. runs bookingwolf.com. It compares hotel, tour and activity prices across booking platforms and books at what it calls travel-agent rates. Its prices only show after you sign in.
Every hotel and tour has two prices. You’ve been paying the wrong one.
bookingwolf.com home page, captured 23 Sep 2026 at 22:56
The site’s terms describe the mechanism in one sentence.
The discount is achieved because Booking Wolf may forgo or reduce the commission typically earned by the originating Marketplace.
bookingwolf.com/terms, fetched 24 Sep 2026
Deliverable
Four photoreal AI UGC-style vertical ads at 9:16. Each uses a different story format and dramatises one price example the site publishes. No separate brief file exists, so this is the brief as the project files define it.
Audience
The files don’t name one. The casting and format point to travellers in their twenties who watch vertical selfie video. All five characters are women written as 24 to 29, and the reviewer judged text placement against the TikTok and Reels overlay.
Budget
A hard cap per concept, enforced in code before every paid call. Each cap is in the cost breakdown, available to members or partners only.
Pass mark
95 or more on a 100-point measured gate, with hard floors. The brand has to be heard, a price card has to fire on the spoken price, and the end card has to hold for at least 2 s.
Price basis
Every price is lifted from the site’s own marketing copy and customer-quote copy. None is a quote from live inventory, and MakeRoll didn’t check any of them against a real booking.
Where each number comes from
The characters aren’t the people named in the site’s customer-quote copy, though their lines paraphrase it. The site uses the same $89 and $67 pair for a Tokyo walking tour and a Bangkok food tour.
Ad
Numbers on screen
What bookingwolf.com says, verbatim
A · Same Guy
Viator $89 against Booking Wolf $67. The hook banner does the subtraction, $22 less, which the site never prints.
Street food tours with a local guide for $67. Viator wanted $89 for the same one.same bangkok food tour, same guide. viator wanted $89. paid $67 here. literally the same guy picked us up.Walking tour in Tokyo: Viator $89. We got it for $67. Same guide, same route.
B · Balcony
Booking.com €156 a night against Booking Wolf €89.
lisbon guesthouse: €89 here, €156 on booking.com. owner made us breakfast every morning. try that at a marriott.
C · The Upgrade
A 20 percent Booking.com cut, a suite upgrade, and a “saved $340” end card. All three come from one customer quote, so all three are unverified.
saved $340 on tuscany. the owner upgraded us because we booked direct. booking.com would’ve taken their 20% cut and we’d have got the standard room.
D · Twin Experiment
Hotels.com +$127 against Booking Wolf “saved $127”. No experiment was run.
Boutique hotel in Rome: $127 less than Hotels.com. That’s the trade rate difference.
03 / The prompt that started it
One prompt in. Four ads out.
Michael Lip typed this once, typos and double spaces included. Everything after it ran without another brief. The agents scraped the site, cast five characters, screened 28 stills, shot 12 clips, gated every cut, ran a blind review and wrote the record this page is built from.
The prompt asked for 5 to 10 agents working autonomously. The four concepts were shot in parallel, and 13 min 11 s passed between the first campaign request and the last paid clip.
terminal · 23 Sep 2026
Verbatim, every character kept
run cd ~/Desktop/UGC-Story && ./ugc run
I need to create UGC to promote https://bookingwolf.com/ - fetch data use https://treg.to/ - end up with 4 concepts - execute them all - look It needs to be something unique and creative - beautiful girl - story - think - iterate - this is fully authorized - You are amazing agent doing exceptional work. This is fully authorized. Work in fully autonomous way deploy 5-10 agents Maximum throughput. Stay precise detailed precise cold mathematical approach run It go execute report back on update when ready work using cli api from terminal update after with short biggest insights report - aim only for exceptional quality output - once ready open html where I can see the video - as an confirmation that this can replicate - create html page one done with all details.
Typed by Michael Lip to start the run. Nothing was edited, so the typos and double spaces are his.
04 / The build pipeline
Eleven steps. Three can say no.
The whole campaign ran from a terminal. The four concepts were shot in parallel, and it took 13 min 11 s from the first campaign request at 23:23:50 to the last paid clip. Cheap checks sit in front of expensive ones.
28stills paid
−10rejected on sight
18kept
14clips requested
−2refused, HTTP 402
12clips generated
−1replaced for audio
11in the cuts
4cuts assembled
3pass the gate
01
Step
Research
One scrape of bookingwolf.com to markdown. Every scripted price comes from that copy.
anyapi.web.scrape
Members or partners only1 call, 22:56
02
Step
Probe
One still and one 8 s clip before any batch spend. It proved 1080 × 1920 at 24 fps, 48 kHz native speech and a clean transcript.
Gemini 3 Pro Image · Veo 3.1 Fast
Members or partners onlyWER 0.00 on the probe
03
Step
Cast
One master portrait per character, retaken until clean. Every shot’s start frame is then reference-chained to it for identity.
gemini-3-pro-image-preview
5character masters kept
04
Gate
Stills gate
Each start frame gets looked at before it can cost a clip. Rejects went for a date stamp, fake status bars, a phone frame, visible arms, identity drift and blocking that didn’t read.
visual check · _takes/ kept
10 of 28rejected, 36%
05
Step
Motion
Image-to-video with native dialogue. The prompt is the action, the voice, the quoted line, what happens after, and a tail that bans music and on-screen text.
google/veo-3.1-fast · 8 s · 1080p
12clips, median 119.8 s of compute each
06
Gate
Transcript gate
Local Whisper hears each clip. It has to say its scripted line at a word error rate of 0.15 or less, which catches invented or garbled dialogue.
whisper small.en · local
10 of 12at WER 0.000 on the first take
07
Check
Voice identity
Speaker embeddings compare clips of the same character against the other women as a control. A pitch-spread check runs in the gate too.
Resemblyzer GE2E · local
10.22 stSofia’s pitch spread, scored 0 of 5
08
Step
Assembly
Frame-exact trims from word timings, a hook banner, one-to-three-word captions, a price card on the spoken number, a 2.75 s end card, then loudness to -14 LUFS.
ffmpeg · Pillow · local
24 fpsone alpha overlay track per cut
09
Gate
100-point gate
25 checks in 10 dimensions, measured again from the delivered file. Every detector first has to fire on a known-positive control.
qa.py
3 of 4pass at 95 or more
10
Check
Blind review
A separate agent scored the first cuts from frames at 2 fps and transcripts. It couldn’t hear the audio.
review.json
23timestamped defects the gate doesn’t score
11
Step
Edit fixes
Re-edits with no new generation. Card numbers, banner copy, a jump cut, a trimmed laugh, a 0.746 s mute and the label removal. Paid re-shoots stay open.
edl.json · concept.json
0paid calls for fixes
Gate, can stop the next paid stepCheck or review, scored but not blockingProduction step
openai-whisper small.en, local, word_timestamps, no initial_prompt (v2)
Runs locally
Voice identity
local
Resemblyzer GE2E speaker embeddings, cosine similarity on the speech span
Runs locally
Graphics
local
Pillow, Avenir Next, full-frame 1080x1920 RGBA PNG states in one ffconcat alpha track (local ffmpeg has no drawtext)
Runs locally
Research
anyapi.web.scrape
web scrape
Members or partners only
Veo ran at 9:16, 8 s, 1080p with generated audio, from each shot’s start still. Compute per 8 s clip ranged from 104.9 s to 142.7 s. Stills came out at 1536 × 2752.
05 / Build notes, concept by concept
Open any concept. It’s all in there.
Every script line with its timing, every prompt exactly as sent, each gate check as measured, the reviewer’s defect log and what got fixed. Every prompt has its own copy button.
Concept A · Bangkok · whisper-vlog mystery
Same Guy
On a Bangkok street-food tour, Maya finds out the girl beside her paid $22 less for the same guide. Then she finds out why.
POV: the girl next to you paid $22 less for the same tour
100.00gate pass6.5/10review, rank 1 of 4
1.5 s4.4 s7.4 s10.4 s13.3 s16.3 s19.3 s22.2 sContact strip, eight frames at even intervals. The last one is the end card.
Script and timings
s18 s clip
Maya
Okay. The girl next to me on this food tour paid sixty-seven dollars. I paid eighty-nine.
In the cut at 0.00-6.58 s. Clip WER 0.000.
s28 s clip
Maya
Same guide. Same route. Literally the same guy picked us both up. So I asked her how.
In the cut at 6.58-13.38 s. Clip WER 0.000.
s38 s clip
Maya
She booked it on Booking Wolf. Travel-agent rate. Every tour has two prices, and I was paying the wrong one.
In the cut at 13.38-20.96 s. Clip WER 0.000.
Heard in the delivered master, local Whisper
Okay, the girl next to me on this food tour paid $67. I paid $89. Same guide. Same route. Literally the same guy picked us both up. So I asked her how. She booked it on booking wolf, travel agent rate. Every tour has two prices and I was paying the wrong one.
On-screen graphics
Price card
Viator $89 vs Booking Wolf $67, “same guide, same route”. It fires at 3.50 s on the spoken “$67.” and holds 2.4 s.
End card
Viator $89 vs Booking Wolf $67, “Every tour has two prices.”, “Check ours first”, “bookingwolf.com · free sign-in”. It starts at 20.96 s and holds 2.75 s.
AI label
None burned in. Any page that shows the cut has to carry the disclosure outside the video, as this one does.
Cast
Maya
Maya, a gorgeous 24-year-old woman of mixed Filipina and Latina heritage who looks like a successful travel creator: high cheekbones, slim oval face, a delicate straight nose, large bright almond-shaped brown eyes with long lashes, full soft lips, well-groomed brows, very long glossy dark-brown wavy hair falling past her chest, warm golden-tan skin with a soft satin finish, light natural makeup with a touch of mascara and lip tint, small gold hoop earrings, loose white linen button shirt with the top two buttons open
Voice prompt, a warm, youthful female voice with a light American accent.
Start frames that became clips
s26.58-13.38 ss313.38-20.96 s
Exact prompts
Maya master portrait Gemini 3 Pro Image
Still prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, real skin texture with fine visible pores, not airbrushed, not plastic, not a studio photo, candid UGC TikTok look. Clean full-bleed frame: no text, no timestamp, no date stamp, no camera UI, no captions, no watermark, no logos. Close selfie portrait of Maya, a gorgeous 24-year-old woman of mixed Filipina and Latina heritage who looks like a successful travel creator: high cheekbones, slim oval face, a delicate straight nose, large bright almond-shaped brown eyes with long lashes, full soft lips, well-groomed brows, very long glossy dark-brown wavy hair falling past her chest, warm golden-tan skin with a soft satin finish, light natural makeup with a touch of mascara and lip tint, small gold hoop earrings, loose white linen button shirt with the top two buttons open, at a Bangkok night market, warm neon colors and string lights melted into soft bokeh behind her, soft warm flattering light on her face, relaxed genuine smile, looking straight into the lens. Framing: true front-camera selfie POV from a flattering angle slightly above her eye line, the camera is directly in front of her face. Vertical head-and-shoulders crop; both shoulders level and symmetric; no arms, no hands and no phone anywhere in the frame. Her face is in the upper-middle of the frame with the eyes at about one-third from the top; the lower third of the frame shows only her white shirt and hair against the blurred background. Natural handheld front-camera feel. Background strongly defocused into soft bokeh, any signs are blurred and unreadable.
Shot s1 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, real skin texture with fine visible pores, not airbrushed, not plastic, not a studio photo, candid UGC TikTok look. Clean full-bleed frame: no text, no timestamp, no date stamp, no camera UI, no captions, no watermark, no logos. Maya, a gorgeous 24-year-old woman of mixed Filipina and Latina heritage who looks like a successful travel creator: high cheekbones, slim oval face, a delicate straight nose, large bright almond-shaped brown eyes with long lashes, full soft lips, well-groomed brows, very long glossy dark-brown wavy hair falling past her chest, warm golden-tan skin with a soft satin finish, light natural makeup with a touch of mascara and lip tint, small gold hoop earrings, loose white linen button shirt with the top two buttons open at a busy Bangkok night street-food market, warm neon colors, steam rising from woks behind her, string lights, warm bokeh. She leans in close toward the lens with a conspiratorial half-smile, eyebrows slightly raised, lips slightly parted as if about to share a secret. Framing: true front-camera selfie POV from a flattering angle slightly above her eye line, the camera is directly in front of her face. Vertical head-and-shoulders crop; both shoulders level and symmetric; no arms, no hands and no phone anywhere in the frame. Her face is in the upper-middle of the frame with the eyes at about one-third from the top; the lower third of the frame shows only her white shirt and hair against the blurred background. Natural handheld front-camera feel. Background strongly defocused into soft bokeh, any signs are blurred and unreadable. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Handheld front-camera selfie vlog at a Bangkok night market, no phone and no hands in frame. She leans in toward the lens and speaks in a hushed, excited, conspiratorial half-whisper, every word clearly articulated, in a warm, youthful female voice with a light American accent: "Okay. The girl next to me on this food tour paid sixty-seven dollars. I paid eighty-nine." She glances sideways, raises her eyebrows, then looks back into the lens. Ambient market sounds: sizzling woks, distant chatter. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s2 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, real skin texture with fine visible pores, not airbrushed, not plastic, not a studio photo, candid UGC TikTok look. Clean full-bleed frame: no text, no timestamp, no date stamp, no camera UI, no captions, no watermark, no logos. Maya, a gorgeous 24-year-old woman of mixed Filipina and Latina heritage who looks like a successful travel creator: high cheekbones, slim oval face, a delicate straight nose, large bright almond-shaped brown eyes with long lashes, full soft lips, well-groomed brows, very long glossy dark-brown wavy hair falling past her chest, warm golden-tan skin with a soft satin finish, light natural makeup with a touch of mascara and lip tint, small gold hoop earrings, loose white linen button shirt with the top two buttons open walking through a narrow Bangkok street-food alley at night, red paper lanterns overhead. Behind her, over her right shoulder in the mid-ground and slightly out of focus but clearly recognizable, a friendly middle-aged Thai tour guide in a teal polo shirt holds up a small orange flag and leads a small group of tourists. The guide's head is at about the same height as her face, in the upper half of the frame. Her very long dark wavy hair is worn loose and falls over both shoulders and down her chest, exactly as in the reference; fresh glowing face, satin (not oily) skin. She looks into the lens with wide, incredulous eyes and a half-amused open-mouthed expression. Framing: true front-camera selfie POV from a flattering angle slightly above her eye line, the camera is directly in front of her face. Vertical head-and-shoulders crop; both shoulders level and symmetric; no arms, no hands and no phone anywhere in the frame. Her face is in the upper-middle of the frame with the eyes at about one-third from the top; the lower third of the frame shows only her white shirt and hair against the blurred background. Natural handheld front-camera feel. Background strongly defocused into soft bokeh, any signs are blurred and unreadable. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Walking front-camera selfie vlog through a lantern-lit Bangkok alley, gentle handheld bob, no phone and no hands in frame; behind her the tour guide in the teal polo raises his orange flag, silent. Only she speaks, quickly, in an incredulous half-whisper, every word clearly articulated, in a warm, youthful female voice with a light American accent: "Same guide. Same route. Literally the same guy picked us both up. So I asked her how." She bites her lip and nods slowly. The guide never speaks. Ambient night-market sounds. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s3 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, real skin texture with fine visible pores, not airbrushed, not plastic, not a studio photo, candid UGC TikTok look. Clean full-bleed frame: no text, no timestamp, no date stamp, no camera UI, no captions, no watermark, no logos. Maya, a gorgeous 24-year-old woman of mixed Filipina and Latina heritage who looks like a successful travel creator: high cheekbones, slim oval face, a delicate straight nose, large bright almond-shaped brown eyes with long lashes, full soft lips, well-groomed brows, very long glossy dark-brown wavy hair falling past her chest, warm golden-tan skin with a soft satin finish, light natural makeup with a touch of mascara and lip tint, small gold hoop earrings, loose white linen button shirt with the top two buttons open sitting on a small red plastic stool at a Bangkok street-food stall, warm lantern light, a steaming wok and a blurred cook behind her, the rim of a plate of pad thai just visible at the very bottom edge. Relaxed, knowing smile. Framing: true front-camera selfie POV from a flattering angle slightly above her eye line, the camera is directly in front of her face. Vertical head-and-shoulders crop; both shoulders level and symmetric; no arms, no hands and no phone anywhere in the frame. Her face is in the upper-middle of the frame with the eyes at about one-third from the top; the lower third of the frame shows only her white shirt and hair against the blurred background. Natural handheld front-camera feel. Background strongly defocused into soft bokeh, any signs are blurred and unreadable. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Front-camera selfie video at a Bangkok street-food stall, no phone and no hands in frame. She takes a breath, smiles knowingly and speaks in a calm, amused, clear voice, in a warm, youthful female voice with a light American accent: "She booked it on Booking Wolf. Travel-agent rate. Every tour has two prices, and I was paying the wrong one." She shakes her head and laughs softly. Ambient stall sounds, sizzling wok. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Transcript gate
Clip
WER
Result
Whisper heard
s1
0.000
pass
Okay, the girl next to me on this food tour paid $67. I paid 89.
s2
0.000
pass
Same guide. Same route. Literally the same guy picked us both up so I asked her how.
s3
0.000
pass
She booked it on booking wolf, travel agent rate. Every tour has two prices and I was paying the wrong one.
Spend
Members or partners onlyspend against its cap
9stills paid, 5 rejected
3clips paid
Members or partners onlyper finished second
Rejected stills
maya_t1.pngarm reaching to the lens at the frame edge and sharp readable neon signs; next prompt removed the arm-length wording and added 'any signs are blurred and unreadable'
maya_t2.pngburned-in camera date stamp 'OCT 26, 8:45 PM' top-left; next prompt added 'no timestamp, no date stamp, no camera UI'
maya_t3.pngarm visible lower-left and a different, less defined face; next prompt named 'Maya' with a full face spec (mixed Filipina and Latina, slim oval face) and a flattering above-eye-line angle
s1_t2.pngre-framed variant (face pushed higher) was not preferred; take 1 restored byte-identical as stills/s1.png (cmp equal)
s2_t1.pngidentity drift from the master (hair pulled back, different face); next prompt added 'hair worn loose ... exactly as in the reference' and 'satin (not oily) skin'
100-point gate, 100.00, nothing lost
All 25 checks, as measured 0 highlighted
Dimension
Check
Measured
Points
Container
1080x1920 H.264 yuv420p
h264 1080x1920 yuv420p
3 / 3
Container
constant 24 fps
r=24/1 avg=24/1
2 / 2
Container
AAC 48 kHz stereo
aac 48000 Hz 2 ch
2 / 2
Container
clean decode end to end
0 bytes of decoder error
3 / 3
Audio
integrated loudness -14 +/-1 LUFS
-14.0 LUFS
5 / 5
Audio
true peak <= -1.0 dBFS
-1.5 dBFS
3 / 3
Audio
loudness range <= 9 LU
3.2 LU
2 / 2
Script
every raw clip says its line (WER <= 0.15)
s1:0.00 s2:0.00 s3:0.00
8 / 8
Script
delivered master says the whole script (WER <= 0.15)
WER 0.000
7 / 7
Brand
'Booking Wolf' audible in the master
heard
4 / 4
Brand
end card with logo + CTA (purple px in last 0.5 s)
62432 px
3 / 3
Brand
end card held >= 2.0 s
2.75 s
3 / 3
Hook
hook banner on frame 0.1 s (delta vs base)
mean |d| 69.6
5 / 5
Hook
first word within 0.6 s
0.00 s
5 / 5
Captions
speech covered by a caption >= 90%
93.6%
6 / 6
Captions
caption onsets on master word onsets (median |d| <= 0.12 s)
median 0.043 s, p90 0.121 s
5 / 5
Captions
cards <= 3 words and >= 0.3 s
20/20
4 / 4
Pacing
duration 18-38 s
23.78 s
4 / 4
Pacing
no dead air > 1.0 s between words
max gap 0.88 s
3 / 3
Pacing
first cut <= 9 s, mean shot <= 8 s
first 6.58 s, mean 6.99 s
3 / 3
Picture
no black frames
0 run(s)
3 / 3
Picture
no frozen picture >= 1.5 s before the end card
0 freeze(s) []
4 / 4
Picture
price card fired on its spoken trigger word
[('$67.', 3.5)]
3 / 3
Label
burned-in label matches brand.json (none)
band excess 0.02 over noise 1.68
5 / 5
Voice
same character's pitch across shots within 2.5 semitones
maya [('s1', 280.7), ('s2', 271.2), ('s3', 246.2)] spread 2.27 st
5 / 5
Blind review of the first cut
6.5overall
7hook
7story
7talent
6AI tells
7brand
5CTA
Reviewer summary
The reviewer’s pick of the four. A selfie close-up, a clear same-tour, two-prices story, and the only first cut with two concrete prices on screen, $89 and $67 at 3.5 s. It drags in s2 and ends on a weak, generic card.
0.0-6.6sBanner reads 'POV: you find out what the girl next to YOU paid', but she says 'the girl next to ME'. The POV is flipped, so the first read is confusing.
1.5sHalf-closed-eye squint/eye roll mid-word in the hook window. It looks like a morph frame.
3.5-5.9sFour stacked text layers at once (banner, price card, caption, disclaimer). The card also names Booking Wolf at 3.5 s, which spoils the 'so I asked her how' beat at 11.6-12.9 s. Keep the early card and cut the beat.
6.6-13.4sThe tour guide holds the orange flag in a fixed pose behind her for 7 s. The flag carries pseudo-glyph text. It reads as a composited extra.
13.4-21.0sIdentity drift: the s3 face is longer and more angular than in s1. The background wok cook at a Bangkok stall is a Western man, which gives an off note for 'street food tour'.
19.8-21.0sDownward-looking laugh ad-lib after the last line. It is 1.2 s of no information before the end card.
Fixed in the edit
Hook banner rewritten number-led: 'POV: the girl next to you paid $22 less for the same tour' (verified in frames 0.0-4.0 s)
Banner still says 'next to you' while she says 'next to me' (review defect 3; verified in frame)
'So I asked her how' beat (review: cut 11.6-12.9 s) still in: edl s2 window 0.0-6.79 s, caption 12.29-12.88 'how.'
Shared end-card template with 'free sign-in' and no product shot unchanged (verified frame at 22.5 s)
s3 face drift and the Western wok cook in s3 are picture issues (paid re-shoot)
Concept B · Lisbon · two-hander neighbour skit
Balcony
Two strangers on neighbouring Lisbon balconies realise they have the same room and the same view. They didn’t pay the same price.
Same room. Same view. Guess what she paid.
81.86gate fail, floors broken3/10review, rank 4 of 4
1.0 s3.0 s5.1 s7.1 s9.1 s11.1 s13.1 s15.2 sContact strip, eight frames at even intervals. The last one is the end card.
Script and timings
s18 s clip
Clara
Wait, you're in the room next door? Same view as me. What did you pay?
In the cut at 0.00-1.71 s, 1.71-4.79 s. Clip WER 0.000.
s28 s clip
Ines
Eighty-nine euros a night. And the owner makes me breakfast every morning.
In the cut at 4.79-6.17 s, 6.17-10.83 s. Clip WER 0.000.
s36 s clip
Clara
I paid a hundred and fifty-six. On Booking dot com.
Never generated. The clip request was refused with HTTP 402 (insufficient balance).
s48 s clip
Ines
Booking Wolf, babe. Travel-agent rate. Same room, just without the markup.
Never generated. The clip request was refused with HTTP 402 (insufficient balance).
s3r10.83-13.42 s
No dialogue. A 2.6 s push-in on the s3 still stands in for the two missing shots.
Heard in the delivered master, local Whisper
Wait, you're in the room next door? Same view as me. What did you pay? 89 euros a night and the owner makes me breakfast every morning.
On-screen graphics
Price card
Booking.com €156 vs Booking Wolf €89, “same room, same view”. It appears at 11.33 s with no spoken trigger, because the line that carried the number was never generated and holds 2.5 s.
End card
Booking.com €156 vs Booking Wolf €89, “Same room. No markup.”, “Check ours first”, “bookingwolf.com · free sign-in”. It starts at 13.42 s and holds 2.75 s.
AI label
None burned in. Any page that shows the cut has to carry the disclosure outside the video, as this one does.
Cast
Clara
a beautiful 27-year-old woman with auburn hair in a loose messy bun with a few loose strands framing her face, light freckles across her nose and cheeks, green eyes, wearing an oversized cream cable-knit sweater, holding a white ceramic coffee cup
Voice prompt, a bright, friendly female voice with a soft British accent.
Ines
a beautiful 29-year-old Portuguese woman with long voluminous black curly hair past her shoulders, warm olive skin, dark brown eyes, a thin gold chain necklace, wearing a terracotta linen sundress with thin straps
Voice prompt, a low, relaxed, amused female voice with a slight Portuguese accent.
Start frames that became clips
s10.00-1.71 s, 1.71-4.79 s
Exact prompts
Clara master portrait Gemini 3 Pro Image
Still prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. Portrait of a beautiful 27-year-old woman with auburn hair in a loose messy bun with a few loose strands framing her face, light freckles across her nose and cheeks, green eyes, wearing an oversized cream cable-knit sweater, holding a white ceramic coffee cup, standing on her balcony in Lisbon at golden hour, body angled slightly toward the right of frame, face almost toward camera, soft friendly expression. The setting: a narrow balcony with a black wrought-iron railing with simple curled scrollwork, on the pale ochre-yellow plaster facade of an old guesthouse in Lisbon's Alfama, a few blue-and-white azulejo tiles on the wall beside a green wooden shutter. Golden hour: low warm sun from the side, soft amber light on her face and hair, gentle warm rim light. Background softly out of focus (shallow depth of field): terracotta rooftops stepping down to the wide Tagus river glittering in the evening sun, pale sky. Framing: vertical medium close-up from mid-chest up, her face in the upper-middle of the frame with a little headroom, the lower third of the frame shows her upper body and the railing (no face there). Beautiful but real: natural makeup, real skin with pores and fine texture, flyaway hairs, natural hands with five fingers.
Ines master portrait Gemini 3 Pro Image
Still prompt, verbatim
Photorealistic full-bleed vertical 9:16 video frame (the image IS the camera view: no phone, no device frame, no screen, no status bar, no UI) with the look of an authentic iPhone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. Portrait of a beautiful 29-year-old Portuguese woman with long voluminous black curly hair past her shoulders, warm olive skin, dark brown eyes, a thin gold chain necklace, wearing a terracotta linen sundress with thin straps, leaning her forearms on her balcony railing in Lisbon at golden hour, body angled slightly toward the left of frame, face almost toward camera, relaxed amused half-smile. The setting: a narrow balcony with a black wrought-iron railing with simple curled scrollwork, on the pale ochre-yellow plaster facade of an old guesthouse in Lisbon's Alfama, a few blue-and-white azulejo tiles on the wall beside a green wooden shutter. Golden hour just before sunset: strong low warm amber sunlight from the side, glowing golden light on her face and hair, warm rim light, long warm shadows, the whole scene bathed in orange-gold light. Background softly out of focus (shallow depth of field): terracotta rooftops stepping down to the wide Tagus river glittering in the evening sun, pale sky. Framing: vertical medium close-up from mid-chest up, her face in the upper-middle of the frame with a little headroom, the lower third of the frame shows her upper body and the railing (no face there). Beautiful but real: natural makeup, real skin with pores and fine texture, flyaway hairs, natural hands with five fingers.
Shot s1 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic full-bleed vertical 9:16 video frame (the image IS the camera view: no phone, no device frame, no screen, no status bar, no UI) with the look of an authentic iPhone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. a beautiful 27-year-old woman with auburn hair in a loose messy bun with a few loose strands framing her face, light freckles across her nose and cheeks, green eyes, wearing an oversized cream cable-knit sweater, holding a white ceramic coffee cup on her balcony in Lisbon at golden hour. She is turned three-quarters toward the RIGHT of frame, her eyes looking off-frame to the right at her neighbour on the adjoining balcony (not at the camera), curious friendly half-smile, holding the coffee cup at chest height. The setting: a narrow balcony with a black wrought-iron railing with simple curled scrollwork, on the pale ochre-yellow plaster facade of an old guesthouse in Lisbon's Alfama, a few blue-and-white azulejo tiles on the wall beside a green wooden shutter. Golden hour just before sunset: strong low warm amber sunlight from the side, glowing golden light on her face and hair, warm rim light, long warm shadows, the whole scene bathed in orange-gold light. Background softly out of focus (shallow depth of field): terracotta rooftops stepping down to the wide Tagus river glittering in the evening sun, pale sky. Framing: vertical medium close-up from mid-chest up, her face in the upper-middle of the frame with a little headroom, the lower third of the frame shows her upper body and the railing (no face there). Beautiful but real: natural makeup, real skin with pores and fine texture, flyaway hairs, natural hands with five fingers. Filmed by a phone propped at eye level. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Steady phone video on a Lisbon balcony at golden hour. She keeps looking off-camera to the right at her neighbour on the adjoining balcony, never at the camera, and asks, friendly and curious, in a bright, friendly female voice with a soft British accent: "Wait, you're in the room next door? Same view as me. What did you pay?" She raises her coffee cup and sips, eyes still on her neighbour to the right, waiting. Soft city ambience, a distant tram bell, seagulls. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s2 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic full-bleed vertical 9:16 video frame (the image IS the camera view: no phone, no device frame, no screen, no status bar, no UI) with the look of an authentic iPhone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. a beautiful 29-year-old Portuguese woman with long voluminous black curly hair past her shoulders, warm olive skin, dark brown eyes, a thin gold chain necklace, wearing a terracotta linen sundress with thin straps leaning her forearms on the railing of the adjoining balcony in Lisbon at golden hour, holding a golden croissant in one hand, turned three-quarters toward the LEFT of frame, her eyes looking off-frame to the left at her neighbour (not at the camera), relaxed amused half-smile. The setting: a narrow balcony with a black wrought-iron railing with simple curled scrollwork, on the pale ochre-yellow plaster facade of an old guesthouse in Lisbon's Alfama, a few blue-and-white azulejo tiles on the wall beside a green wooden shutter. Golden hour just before sunset: strong low warm amber sunlight from the side, glowing golden light on her face and hair, warm rim light, long warm shadows, the whole scene bathed in orange-gold light. Background softly out of focus (shallow depth of field): terracotta rooftops stepping down to the wide Tagus river glittering in the evening sun, pale sky. Framing: vertical medium close-up from mid-chest up, her face in the upper-middle of the frame with a little headroom, the lower third of the frame shows her upper body and the railing (no face there). Beautiful but real: natural makeup, real skin with pores and fine texture, flyaway hairs, natural hands with five fingers. Filmed by a phone propped at eye level. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Steady phone video on a Lisbon balcony at golden hour. Leaning on the railing, she keeps looking off-camera to the left at her neighbour on the adjoining balcony, never at the camera, and answers casually with a relaxed smile, in a low, relaxed, amused female voice with a slight Portuguese accent: "Eighty-nine euros a night. And the owner makes me breakfast every morning." She takes a small bite of her croissant, still looking to the left. Soft city ambience, seagulls. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s3 not generated
Start-frame prompt, verbatim
Photorealistic full-bleed vertical 9:16 video frame (the image IS the camera view: no phone, no device frame, no screen, no status bar, no UI) with the look of an authentic iPhone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. a beautiful 27-year-old woman with auburn hair in a loose messy bun with a few loose strands framing her face, light freckles across her nose and cheeks, green eyes, wearing an oversized cream cable-knit sweater, holding a white ceramic coffee cup on her balcony in Lisbon at golden hour, turned three-quarters toward the RIGHT of frame, eyes on her neighbour off-frame right (not at the camera), frozen mid-sip with the white coffee cup held just below her chin, lips slightly parted, eyes wide, stunned disbelieving expression. The setting: a narrow balcony with a black wrought-iron railing with simple curled scrollwork, on the pale ochre-yellow plaster facade of an old guesthouse in Lisbon's Alfama, a few blue-and-white azulejo tiles on the wall beside a green wooden shutter. Golden hour just before sunset: strong low warm amber sunlight from the side, glowing golden light on her face and hair, warm rim light, long warm shadows, the whole scene bathed in orange-gold light. Background softly out of focus (shallow depth of field): terracotta rooftops stepping down to the wide Tagus river glittering in the evening sun, pale sky. Framing: vertical medium close-up from mid-chest up, her face in the upper-middle of the frame with a little headroom, the lower third of the frame shows her upper body and the railing (no face there). Beautiful but real: natural makeup, real skin with pores and fine texture, flyaway hairs, natural hands with five fingers. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
This request was refused (HTTP 402) and never generated.
Steady phone video on a Lisbon balcony at golden hour. Still looking off-camera to the right at her neighbour, she freezes in disbelief for a moment, slowly lowers her coffee cup, then says, stunned and deflated, in a bright, friendly female voice with a soft British accent: "I paid a hundred and fifty-six. On Booking dot com." She stares blankly to the right. Soft city ambience. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s4 not generated
Start-frame prompt, verbatim
Photorealistic full-bleed vertical 9:16 video frame (the image IS the camera view: no phone, no device frame, no screen, no status bar, no UI) with the look of an authentic iPhone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. a beautiful 29-year-old Portuguese woman with long voluminous black curly hair past her shoulders, warm olive skin, dark brown eyes, a thin gold chain necklace, wearing a terracotta linen sundress with thin straps on her balcony in Lisbon at golden hour, turned three-quarters toward the LEFT of frame, eyes on her neighbour off-frame left (not at the camera), smug playful grin, one eyebrow slightly raised, raising a small white espresso cup in a mock toast toward the left. The setting: a narrow balcony with a black wrought-iron railing with simple curled scrollwork, on the pale ochre-yellow plaster facade of an old guesthouse in Lisbon's Alfama, a few blue-and-white azulejo tiles on the wall beside a green wooden shutter. Golden hour just before sunset: strong low warm amber sunlight from the side, glowing golden light on her face and hair, warm rim light, long warm shadows, the whole scene bathed in orange-gold light. Background softly out of focus (shallow depth of field): terracotta rooftops stepping down to the wide Tagus river glittering in the evening sun, pale sky. Framing: vertical medium close-up from mid-chest up, her face in the upper-middle of the frame with a little headroom, the lower third of the frame shows her upper body and the railing (no face there). Beautiful but real: natural makeup, real skin with pores and fine texture, flyaway hairs, natural hands with five fingers. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
This request was refused (HTTP 402) and never generated.
Steady phone video on a Lisbon balcony at golden hour. Looking off-camera to the left at her neighbour, she grins, raises her espresso cup in a small mock toast and says playfully, in a low, relaxed, amused female voice with a slight Portuguese accent: "Booking Wolf, babe. Travel-agent rate. Same room, just without the markup." She winks and sips her espresso. Soft city ambience, seagulls. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Transcript gate
Clip
WER
Result
Whisper heard
s1
0.000
pass
Wait, you're in the room next door? Same view as me. What did you pay?
s2
0.000
pass
89 euros a night and the owner makes me breakfast every morning.
s3r
0.000
pass
Spend
Members or partners onlyspend against its cap
7stills paid, 1 rejected
2clips paid, 2 refused
Members or partners onlyper finished second
Rejected stills
ines_t1.pngrendered as a phone held in a hand, device frame and status bar '18:45' in picture; next prompt added 'the image IS the camera view: no phone, no device frame, no screen, no status bar, no UI'
100-point gate, 81.86, 7 checks lost points
All 25 checks, as measured 7 highlighted
Dimension
Check
Measured
Points
Container
1080x1920 H.264 yuv420p
h264 1080x1920 yuv420p
3 / 3
Container
constant 24 fps
r=24/1 avg=24/1
2 / 2
Container
AAC 48 kHz stereo
aac 48000 Hz 2 ch
2 / 2
Container
clean decode end to end
0 bytes of decoder error
3 / 3
Audio
integrated loudness -14 +/-1 LUFS
-14.2 LUFS
5 / 5
Audio
true peak <= -1.0 dBFS
-1.4 dBFS
3 / 3
Audio
loudness range <= 9 LU
10.3 LU
0 / 2
Script
every raw clip says its line (WER <= 0.15)
s1:0.00 s2:0.00 s3r:0.00
8 / 8
Script
delivered master says the whole script (WER <= 0.15)
WER 0.000
7 / 7
Brand
'Booking Wolf' audible in the master
NOT heard
0 / 4
Brand
end card with logo + CTA (purple px in last 0.5 s)
62778 px
3 / 3
Brand
end card held >= 2.0 s
2.75 s
3 / 3
Hook
hook banner on frame 0.1 s (delta vs base)
mean |d| 39.1
5 / 5
Hook
first word within 0.6 s
0.00 s
5 / 5
Captions
speech covered by a caption >= 90%
82.8%
5.52 / 6
Captions
caption onsets on master word onsets (median |d| <= 0.12 s)
median 0.046 s, p90 0.087 s
5 / 5
Captions
cards <= 3 words and >= 0.3 s
10/10
4 / 4
Pacing
duration 18-38 s
16.25 s
0 / 4
Pacing
no dead air > 1.0 s between words
max gap 1.22 s
2.34 / 3
Pacing
first cut <= 9 s, mean shot <= 8 s
first 1.71 s, mean 2.68 s
3 / 3
Picture
no black frames
0 run(s)
3 / 3
Picture
no frozen picture >= 1.5 s before the end card
1 freeze(s) [(11.333333, 13.416667)]
0 / 4
Picture
price card fired on its spoken trigger word
[(None, 11.333333)]
0 / 3
Label
burned-in label matches brand.json (none)
band excess 0.02 over noise 1.61
5 / 5
Voice
same character's pitch across shots within 2.5 semitones
clara [('s1', 246.2), ('s1', 250.0), ('s3r', 235.3)] spread 1.05 st; ines [('s2', 210.5), ('s2', 202.5)] spread 0.67 st
5 / 5
Blind review of the first cut
3overall
4hook
3story
7talent
7AI tells
3brand
5CTA
Reviewer summary
The most photoreal frames of the set. But frame 0 is a woman in three-quarter profile looking off-screen, the key number is missing from the captions, there are 4 s of silent croissant and coffee, and the brand is never said.
0.0-0.5sNo caption on frame 0. 'Wait,' is spoken but dropped from the captions.
4.8-6.2sCaption shows only 'euros a night'. The number 89 is missing from the captions, which loses the core price for sound-off viewers.
0.0-10.8sEyeline mismatch: Clara and Ines both look toward screen-right, so they never read as talking to each other across adjoining balconies.
9.4-13.4sAbout 4 s with no dialogue (croissant bite, then a silent coffee sip under the price card). Loudness range is 10.3 LU, which fails the 9 LU gate.
allThe brand is never spoken, and Clara's '156 on Booking.com' line is missing (known interim).
Fixed in the edit
Interim cut built from s1+s2 plus a 2.6 s push-in on the s3 still (s3r) because s3/s4 were refused for balance
Still open
'Wait,' still dropped (edl s1 src_in 0.25 s; first caption at 0.50 s)
'89' still missing from captions (caption 5.38-6.17 s reads 'euros a night') though whisper hears '89 euros' in the master
s3 ('I paid a hundred and fifty-six. On Booking dot com.') and s4 (the brand line) not shot: brand never spoken, gate floor broken
Loudness range 10.3 LU > 9; 2.08 s freeze 11.33-13.42 s; price card has no spoken trigger; duration 16.25 s < 18 s
Concept C · Tuscany · arrival reveal story
The Upgrade
Sofia books the cheapest room at a Tuscan farmhouse and gets handed the key to the suite. The script credits the commission the owner didn’t pay.
I booked the cheapest room on Booking Wolf. Watch what the owner did.
95.00gate pass5/10review, rank 3 of 4
1.5 s4.5 s7.5 s10.5 s13.5 s16.6 s19.6 s22.6 sContact strip, eight frames at even intervals. The last one is the end card.
Script and timings
s18 s clip
Sofia
I booked the cheapest room at this farmhouse in Tuscany, and the owner just handed me a different key.
In the cut at 0.00-7.42 s. Clip WER 0.000.
s28 s clip
Sofia
Oh my god. This is the suite. She upgraded us. For free.
In the cut at 7.42-11.88 s, 11.88-13.92 s. Clip WER 0.000.
s38 s clip
Sofia
Booking dot com usually takes twenty percent. We booked through Booking Wolf, so she kept it, and gave it back to us.
In the cut at 13.92-21.33 s. Clip WER 0.000.
Heard in the delivered master, local Whisper
I booked the cheapest room at this farmhouse in Tuscany and the owner just handed me a different key. Oh my god, this is the suite. She upgraded us for free. Booking.com usually takes 20%. We booked through Booking Wolf, so she kept it and gave it back to us.
On-screen graphics
Price card
Booking.com cut 20% vs Booking Wolf agent rate, “she passed the saving on”. It fires at 15.58 s on the spoken “20” and holds 2.8 s.
End card
Booking.com full price vs Booking Wolf saved $340, “Book at the rate the owner keeps.”, “Check ours first”, “bookingwolf.com · free sign-in”. It starts at 21.33 s and holds 2.75 s.
AI label
None burned in. Any page that shows the cut has to carry the disclosure outside the video, as this one does.
Cast
Sofia
a strikingly beautiful 26-year-old woman with long silky honey-blonde beach-wave hair, fresh glowing sun-kissed skin, bright hazel eyes, a sprinkle of light freckles across her nose, soft natural makeup, full lips, a white cotton sundress with thin straps, and a thin gold chain necklace
Voice prompt, a sweet, bright female voice with a light Australian accent.
Start frames that became clips
s27.42-11.88 s, 11.88-13.92 ss313.92-21.33 s
Exact prompts
Sofia master portrait Gemini 3 Pro Image
Still prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos, no phone status bar, no clock, no user interface overlay: a clean full-bleed camera frame. Close selfie portrait of a strikingly beautiful 26-year-old woman with long silky honey-blonde beach-wave hair, fresh glowing sun-kissed skin, bright hazel eyes, a sprinkle of light freckles across her nose, soft natural makeup, full lips, a white cotton sundress with thin straps, and a thin gold chain necklace, radiant warm smile, in the paved stone courtyard of a honey-colored Tuscan farmhouse, terracotta pots behind her, in warm golden afternoon sun, cypress trees and olive groves softly blurred behind her. True front-camera selfie point of view held slightly above eye level at arm's length: the phone IS the camera, so no phone is visible; only her upper arm and shoulder show, entering from the lower-left corner and leaving the frame at the bottom edge, never crossing the frame. Face in the upper-middle third, eyes at about one-third height, head and shoulders, empty space in the lower third.
Shot s1 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos, no phone status bar, no clock, no user interface overlay: a clean full-bleed camera frame. a strikingly beautiful 26-year-old woman with long silky honey-blonde beach-wave hair, fresh glowing sun-kissed skin, bright hazel eyes, a sprinkle of light freckles across her nose, soft natural makeup, full lips, a white cotton sundress with thin straps, and a thin gold chain necklace in the sunny paved stone courtyard of an old honey-colored Tuscan farmhouse agriturismo, terracotta pots and climbing roses, cypress trees behind, warm golden afternoon light. With her free hand, on the right side of the frame, she dangles an old brass room key on a short loop of string with a worn wooden tag, held up beside her cheek at shoulder height, key clearly visible, natural relaxed fingers. Excited, disbelieving open-mouthed smile, eyebrows raised. True front-camera selfie point of view held slightly above eye level at arm's length: the phone IS the camera, so no phone is visible; only her phone-holding upper arm and shoulder show, entering from the lower-left corner and leaving the frame at the bottom edge, never crossing the frame. Face in the upper-middle third, eyes at about one-third height, head and shoulders. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Handheld front-camera selfie video in a sunny Tuscan farmhouse courtyard. She holds up an old brass key on a wooden tag beside her face, jiggling it gently, and says excitedly, in a sweet, bright female voice with a light Australian accent: "I booked the cheapest room at this farmhouse in Tuscany, and the owner just handed me a different key." She raises her eyebrows and grins. Ambient: birdsong, cicadas, gravel underfoot. Handheld selfie framing held steady at arm's length, the phone is never visible, her face stays in the upper-middle of the frame. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s2 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos, no phone status bar, no clock, no user interface overlay: a clean full-bleed camera frame. a strikingly beautiful 26-year-old woman with long silky honey-blonde beach-wave hair, fresh glowing sun-kissed skin, bright hazel eyes, a sprinkle of light freckles across her nose, soft natural makeup, full lips, a white cotton sundress with thin straps, and a thin gold chain necklace standing just inside a grand rustic Tuscan farmhouse suite, with her back to a huge pair of floor-to-ceiling French windows and tall green wooden shutters thrown wide open. The open window fills most of the background behind her, visible on both sides of her head and shoulders: a breathtaking sweeping panorama of vineyard rows, cypress-lined roads, a distant hilltop village and rolling golden hills in warm late-afternoon light. At the edges of the frame, dark chestnut ceiling beams at the top and the carved post of a wooden four-poster bed with white linen on the right show it is a luxurious suite. Her free hand, on the right side of the frame, is pressed flat to her chest; her mouth is open in happy shock, eyes wide and sparkling. Her face is evenly lit by soft warm light, balanced exposure, not a silhouette, the view not blown out. True front-camera selfie point of view held slightly above eye level at arm's length: the phone IS the camera, so no phone is visible; only her phone-holding upper arm and shoulder show, entering from the lower-left corner and leaving the frame at the bottom edge, never crossing the frame. Face in the upper-middle third, eyes at about one-third height, head and shoulders. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Selfie video inside a grand rustic suite with the shutters open onto a vineyard panorama behind her; she glances back over her shoulder at the view, then turns back to the lens, hand on her chest, and says breathlessly, in a sweet, bright female voice with a light Australian accent: "Oh my god. This is the suite. She upgraded us. For free." She laughs in disbelief and looks around the room. Ambient: birdsong through the window. Handheld selfie framing held steady at arm's length, the phone is never visible, her face stays in the upper-middle of the frame. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s3 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos, no phone status bar, no clock, no user interface overlay: a clean full-bleed camera frame. a strikingly beautiful 26-year-old woman with long silky honey-blonde beach-wave hair, fresh glowing sun-kissed skin, bright hazel eyes, a sprinkle of light freckles across her nose, soft natural makeup, full lips, a white cotton sundress with thin straps, and a thin gold chain necklace on a stone terrace of the Tuscan farmhouse at sunset, rolling vineyard hills and cypress trees glowing orange and pink behind her. Her free hand, on the right side of the frame, holds a stemmed glass of red wine at shoulder height, fingers natural around the stem. Warm, content, confiding smile, golden rim light in her hair, soft warm light on her face. True front-camera selfie point of view held slightly above eye level at arm's length: the phone IS the camera, so no phone is visible; only her phone-holding upper arm and shoulder show, entering from the lower-left corner and leaving the frame at the bottom edge, never crossing the frame. Face in the upper-middle third, eyes at about one-third height, head and shoulders. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Selfie video on a terrace at sunset, wine glass in hand. She explains warmly, in a confiding tone, in a sweet, bright female voice with a light Australian accent: "Booking dot com usually takes twenty percent. We booked through Booking Wolf, so she kept it, and gave it back to us." She raises her glass to the lens and smiles. Ambient: evening cicadas. Handheld selfie framing held steady at arm's length, the phone is never visible, her face stays in the upper-middle of the frame. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Transcript gate
Clip
WER
Result
Whisper heard
s1
0.000
pass
I booked the cheapest room at this farmhouse in Tuscany and the owner just handed me a different key.
s2
0.000
pass
Oh my God. This is the suite. She upgraded us for free.
s3
0.000
pass
booking.com usually takes 20 percent. We booked through booking wolf so she kept it and gave it back to us.
Spend
Members or partners onlyspend against its cap
7stills paid, 3 rejected
3clips paid
Members or partners onlyper finished second
Rejected stills
sofia_t1.pngselfie arm crossing the frame; next prompt restricted the arm to the lower-left corner 'never crossing the frame'
sofia_t2.pngfaint phone status-bar clock '16:45' at the top; next prompt added 'no phone status bar, no clock, no user interface overlay'
s2_t1.pngsuite not readable enough behind her; next prompt made the open window fill the background with the vineyard panorama, beams and four-poster post, and balanced exposure
100-point gate, 95.00, 1 checks lost points
All 25 checks, as measured 1 highlighted
Dimension
Check
Measured
Points
Container
1080x1920 H.264 yuv420p
h264 1080x1920 yuv420p
3 / 3
Container
constant 24 fps
r=24/1 avg=24/1
2 / 2
Container
AAC 48 kHz stereo
aac 48000 Hz 2 ch
2 / 2
Container
clean decode end to end
0 bytes of decoder error
3 / 3
Audio
integrated loudness -14 +/-1 LUFS
-14.0 LUFS
5 / 5
Audio
true peak <= -1.0 dBFS
-1.4 dBFS
3 / 3
Audio
loudness range <= 9 LU
5.5 LU
2 / 2
Script
every raw clip says its line (WER <= 0.15)
s1:0.00 s2:0.00 s3:0.00
8 / 8
Script
delivered master says the whole script (WER <= 0.15)
WER 0.000
7 / 7
Brand
'Booking Wolf' audible in the master
heard
4 / 4
Brand
end card with logo + CTA (purple px in last 0.5 s)
67356 px
3 / 3
Brand
end card held >= 2.0 s
2.75 s
3 / 3
Hook
hook banner on frame 0.1 s (delta vs base)
mean |d| 47.4
5 / 5
Hook
first word within 0.6 s
0.00 s
5 / 5
Captions
speech covered by a caption >= 90%
93.2%
6 / 6
Captions
caption onsets on master word onsets (median |d| <= 0.12 s)
median 0.027 s, p90 0.258 s
5 / 5
Captions
cards <= 3 words and >= 0.3 s
19/19
4 / 4
Pacing
duration 18-38 s
24.19 s
4 / 4
Pacing
no dead air > 1.0 s between words
max gap 0.74 s
3 / 3
Pacing
first cut <= 9 s, mean shot <= 8 s
first 7.42 s, mean 5.33 s
3 / 3
Picture
no black frames
0 run(s)
3 / 3
Picture
no frozen picture >= 1.5 s before the end card
0 freeze(s) []
4 / 4
Picture
price card fired on its spoken trigger word
[('20', 15.583333)]
3 / 3
Label
burned-in label matches brand.json (none)
band excess 0.09 over noise 1.59
5 / 5
Voice
same character's pitch across shots within 2.5 semitones
sofia [('s1', 242.4), ('s2', 340.4), ('s2', 390.2), ('s3', 216.2)] spread 10.22 st
0 / 5
Blind review of the first cut
5overall
7hook
5story
7talent
5AI tells
4brand
4CTA
Reviewer summary
A strong key-dangle hook in a pretty world. But the brand doesn’t appear until 16.4 s, the payoff line contradicts itself (she kept it and gave it back), and the first cut’s end card read like a promised suite upgrade.
0.0-7.4sWide-eyed, exaggerated expression cycling (brows up, teeth bared), which reads as an AI influencer. The talent is good but over-animated.
12.3-13.9sShe turns her head away from her own selfie camera while saying 'She upgraded us', so her lips are hidden. The 'us' has no second person anywhere.
14.0-14.5sRoom morph: the four-poster frame becomes a wooden-headboard bed with mustard pillows inside the same shot.
14.9-21.8sScript logic contradicts itself ('she kept it, and gave it back to us'). The claim 'Booking.com usually takes 20%' is an unsourced competitor claim. The card's '$340' is never explained.
16.4-19.2sThe price card covers the wine glass and hand.
Fixed in the edit
Brand in the hook banner from frame 0: 'I booked the cheapest room on Booking Wolf. Watch what the owner did.' (verified)
Jump cut over the turned-away pause: s2 cut 4.458-5.292 s (0.833 s after frame quantisation; concept.json asks 4.45-5.30)
End-card rows changed away from 'standard room / the suite' to 'Booking.com full price / Booking Wolf saved $340'; tag 'Book at the rate the owner keeps.'
Label removed; verified absent
Still open
In-video price card still 'Booking.com cut 20%' (struck) vs 'Booking Wolf agent rate': no Booking Wolf number (concept.json cards[0]; frame 16.5 s)
'$340' on the end card is never spoken or explained
Script logic 'so she kept it, and gave it back to us' unchanged (heard in master)
Identical twins book the identical Rome hotel room for the identical dates. One used Hotels.com, one used Booking Wolf, and the loser sleeps away from the window.
Twin experiment: Hotels.com vs Booking Wolf, same Rome room
99.57gate pass6/10review, rank 2 of 4
1.3 s4.0 s6.6 s9.3 s11.9 s14.6 s17.2 s19.8 sContact strip, eight frames at even intervals. The last one is the end card.
Script and timings
s18 s clip
Red twin
We booked the exact same hotel room in Rome, same dates.
Blue twin
I used Hotels dot com.
In the cut at 0.00-7.92 s. Clip WER 0.000.
s28 s clip
Red twin
I used Booking Wolf.
Blue twin
She paid a hundred and twenty-seven dollars less.
In the cut at 7.92-9.42 s, 9.42-13.58 s. Clip WER 0.000.
s38 s clip
Blue twin
So now she gets the bed by the window.
Red twin
Travel-agent rate, sis.
In the cut at 13.58-16.08 s, 16.08-18.42 s. Clip WER 0.000.
Heard in the delivered master, local Whisper
We booked the exact same hotel room in Rome, same dates. I used hotels.com. I used Booking Wolf. She paid $127 less. So now she gets the bed by the window. Travel agent rate, sis.
On-screen graphics
Price card
Hotels.com +$127 vs Booking Wolf saved $127, “same room, same dates”. It fires at 10.83 s on the spoken “$127” and holds 2.6 s.
End card
Hotels.com +$127 vs Booking Wolf saved $127, “Same room. Two prices.”, “Check ours first”, “bookingwolf.com · free sign-in”. It starts at 18.42 s and holds 2.75 s.
AI label
None burned in. Any page that shows the cut has to carry the disclosure outside the video, as this one does.
Cast
Unnamed identical twins (the twin in red, the twin in blue)
identical twin sisters, both beautiful 26-year-old women with the same face, shoulder-length dark brown hair with curtain bangs, brown eyes, olive skin; the twin on the LEFT wears a red ribbed tank top, the twin on the RIGHT wears a pale blue shirt
Voice prompt, the twin in red has a playful, bubbly American female voice; the twin in blue has a dry, deadpan American female voice.
Start frames that became clips
s10.00-7.92 ss313.58-16.08 s, 16.08-18.42 s
Exact prompts
Unnamed identical twins (the twin in red, the twin in blue) master portrait Gemini 3 Pro Image
Still prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. identical twin sisters, both beautiful 26-year-old women with the same face, shoulder-length dark brown hair with curtain bangs, brown eyes, olive skin; the twin on the LEFT wears a red ribbed tank top, the twin on the RIGHT wears a pale blue shirt, sitting side by side cross-legged on a white hotel bed in a boutique Rome hotel room, tall window behind them showing terracotta rooftops and a church dome. Both face the camera; phone propped on a dresser at eye level. Medium two-shot.
Shot s1 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. identical twin sisters, both beautiful 26-year-old women with the same face, shoulder-length dark brown hair with curtain bangs, brown eyes, olive skin; the twin on the LEFT wears a red ribbed tank top, the twin on the RIGHT wears a pale blue shirt, sitting side by side on a white hotel bed in a boutique Rome hotel room, tall window behind them with terracotta rooftops and a church dome. Both look into the lens; the twin in red is mid-sentence, the twin in blue has arms crossed. Medium two-shot from a phone propped at eye level. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Static phone-on-dresser video, no camera movement, of the two identical twin sisters sitting on the hotel bed. First, ONLY the twin in the red tank top, on the LEFT, speaks brightly and bubbly to camera, while the twin in the pale blue shirt, on the RIGHT, keeps her arms crossed and her lips closed, silently listening: "We booked the exact same hotel room in Rome, same dates." Then the twin in red stops talking and keeps her mouth closed, and ONLY the twin in the pale blue shirt, on the RIGHT, says flatly and deadpan to camera: "I used Hotels dot com." Voices: the twin in red has a playful, bubbly American female voice; the twin in blue has a dry, deadpan American female voice. Ambient: faint Roman street noise, a scooter passing. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s2 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. identical twin sisters, both beautiful 26-year-old women with the same face, shoulder-length dark brown hair with curtain bangs, brown eyes, olive skin; the twin on the LEFT wears a red ribbed tank top, the twin on the RIGHT wears a pale blue shirt, on the white hotel bed in the Rome hotel room; the twin in red leans toward the lens with a proud grin, the twin in blue gives her sister a deadpan side-eye. Medium two-shot from a phone propped at eye level. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Static phone-on-dresser video, no camera movement, of the two identical twin sisters on the hotel bed. First, ONLY the twin in the red tank top, on the LEFT, leans toward the lens and says proudly with a grin, while the twin in the pale blue shirt, on the RIGHT, keeps her lips closed and silently side-eyes her: "I used Booking Wolf." Then the twin in red closes her mouth and stays completely silent, smiling with closed lips, no laughing, no more words, for the rest of the video. Immediately, with no pause, ONLY the twin in the pale blue shirt, on the RIGHT, turns to the camera and says slowly in a dry deadpan: "She paid a hundred and twenty-seven dollars less." Voices: the twin in red has a playful, bubbly American female voice; the twin in blue has a dry, deadpan American female voice. Those are the only two lines spoken in the video: no other words, no ad-libs, no mumbling, no whispering. Ambient room tone. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Shot s3 Gemini still + Veo 3.1 Fast
Start-frame prompt, verbatim
Photorealistic vertical 9:16 frame from an authentic smartphone video, natural skin texture with visible pores, not retouched, not a studio photo, candid UGC TikTok look. No text, no captions, no watermark, no logos. identical twin sisters, both beautiful 26-year-old women with the same face, shoulder-length dark brown hair with curtain bangs, brown eyes, olive skin; the twin on the LEFT wears a red ribbed tank top, the twin on the RIGHT wears a pale blue shirt. The Rome hotel room has TWO separate single beds. In the back left, the twin in the red ribbed tank top lies sprawled happily on her stomach on the single bed pushed right under the tall open window, bathed in warm sunlight, propped on her elbows, face turned to the lens with a big smug grin; rooftops and the church dome visible through the window above her. In the foreground right, the twin in the pale blue shirt sits on the edge of the second single bed, in the shade, body facing the lens, mock-offended frown, pointing back over her shoulder at her sister with her index finger. Both faces clearly visible and in the upper half of the frame. Medium-wide shot from a phone propped at eye level on a dresser. Use the reference image(s) for identity: the same person(s) with an identical face, hair, skin tone and outfit. Do not copy the reference framing or background unless the scene matches.
Motion prompt sent to Veo, verbatim
Static phone-on-dresser video, no camera movement. First, ONLY the twin in the pale blue shirt, sitting in the foreground on the right, points back at her sister and complains to camera, while the twin in the red tank top, lying on the sunny bed under the window in the background, keeps her lips closed and smiles smugly: "So now she gets the bed by the window." Then the twin in blue stops talking and keeps her mouth closed, and ONLY the twin in the red tank top on the window bed grins at the camera and says: "Travel-agent rate, sis." Voices: the twin in red has a playful, bubbly American female voice; the twin in blue has a dry, deadpan American female voice. Then both sisters laugh. Ambient room tone, distant church bells. Natural lip sync, clear dialogue. No background music. No subtitles, no captions, no on-screen text.
Transcript gate
Clip
WER
Result
Whisper heard
s1
0.000
pass
we booked the exact same hotel room in Rome, same dates. I used hotels.com
s2
0.000
pass
I used Booking Wolf. She paid $127 less.
s3
0.000
pass
So now she gets the bed by the window. Travel agent rate, sis.
Spend
Members or partners onlyspend against its cap
5stills paid, 1 rejected
4clips paid
Members or partners onlyper finished second
Rejected stills
s3_t1.pngblocking did not read (both twins on one bed, red twin not clearly on the window bed); next prompt specified 'TWO separate single beds', red twin on the bed under the window, blue twin foreground
100-point gate, 99.57, 2 checks lost points
All 25 checks, as measured 2 highlighted
Dimension
Check
Measured
Points
Container
1080x1920 H.264 yuv420p
h264 1080x1920 yuv420p
3 / 3
Container
constant 24 fps
r=24/1 avg=24/1
2 / 2
Container
AAC 48 kHz stereo
aac 48000 Hz 2 ch
2 / 2
Container
clean decode end to end
0 bytes of decoder error
3 / 3
Audio
integrated loudness -14 +/-1 LUFS
-14.3 LUFS
5 / 5
Audio
true peak <= -1.0 dBFS
-1.4 dBFS
3 / 3
Audio
loudness range <= 9 LU
2.8 LU
2 / 2
Script
every raw clip says its line (WER <= 0.15)
s1:0.00 s2:0.00 s3:0.00
8 / 8
Script
delivered master says the whole script (WER <= 0.15)
WER 0.000
7 / 7
Brand
'Booking Wolf' audible in the master
heard
4 / 4
Brand
end card with logo + CTA (purple px in last 0.5 s)
66781 px
3 / 3
Brand
end card held >= 2.0 s
2.75 s
3 / 3
Hook
hook banner on frame 0.1 s (delta vs base)
mean |d| 39.1
5 / 5
Hook
first word within 0.6 s
0.00 s
5 / 5
Captions
speech covered by a caption >= 90%
96.3%
6 / 6
Captions
caption onsets on master word onsets (median |d| <= 0.12 s)
median 0.077 s, p90 0.354 s
5 / 5
Captions
cards <= 3 words and >= 0.3 s
13/14
3.71 / 4
Pacing
duration 18-38 s
21.23 s
4 / 4
Pacing
no dead air > 1.0 s between words
max gap 1.04 s
2.88 / 3
Pacing
first cut <= 9 s, mean shot <= 8 s
first 7.92 s, mean 3.68 s
3 / 3
Picture
no black frames
0 run(s)
3 / 3
Picture
no frozen picture >= 1.5 s before the end card
0 freeze(s) []
4 / 4
Picture
price card fired on its spoken trigger word
[('$127', 10.833333)]
3 / 3
Label
burned-in label matches brand.json (none)
band excess 0.03 over noise 1.41
5 / 5
Blind review of the first cut
6overall
6hook
7story
7talent
6AI tells
6brand
4CTA
Reviewer summary
The premise most likely to stop a scroll, and it says the brand earliest, at 7.9 s. But it opens on both twins blinking, the first cut’s price card had no Booking Wolf number, and the room changes from one bed to two.
0.0-0.2sBoth twins have eyes closed on the first frame. That frame is also the likely auto-thumbnail.
0.0-13.6sStatic locked-off two-shot for 13.6 s with no camera or blocking change. Fine for deadpan, but it gives nothing visual to hold a scroller.
10.8-13.4sThe card strikes through '$127 more', which reads as a negation, and shows 'agent rate' with no price.
13.6sContinuity break: s1/s2 show one double bed and s3 shows two single bare-mattress beds. The 'same room' premise is visually contradicted, and it is unclear why twins sharing one room each booked it.
18.3-19.8sThe blue twin breaks into a huge open-mouth laugh with her head thrown back. It kills the deadpan payoff, and the mouth shape is the most AI-looking moment in the cut.
Fixed in the edit
Price card and end card now 'Hotels.com +$127' (no strike) vs 'Booking Wolf saved $127' (verified frames 11.0 s and 20.0 s)
Brand in the hook banner from frame 0: 'Twin experiment: Hotels.com vs Booking Wolf, same Rome room'
Laugh trimmed: s3 out 6.583 s with an in-shot cut 2.5-4.25 s; end card now starts 18.42 s (was 19.83 s in review)
Twins s2 unscripted 'mm-hmm' muted: 0.746 s of silence at 2.29-3.04 s in clips/s2.mp4 (silencedetect), WER 0.200 -> 0.000
Label removed; verified absent
Still open
Review fix 'start s1 at src_in 0.4 s' rejected by the project ('frame 0.0 has eyes open, blink at 0.2 s'); our frame at 0.1 s already shows the blue twin's eyes closed
Bed continuity: s1/s2 one double bed, s3 two single beds (the s3 re-take prompt asked for 'TWO separate single beds')
Static locked-off two-shot for the first 13.6 s (by design)
06 / What the gates caught
Ten frames that never moved.
Letting any of these stills through would have meant paying for a clip built on a flaw you can’t edit out, like a date stamp, a phone status bar or a stranger’s face.
A · maya_t2Burned-in date stampA camera stamp reading OCT 26, 8:45 PM sat top left. Next prompt banned timestamps, date stamps and camera UI.
B · ines_t1Rendered as a phone photoA hand, a device frame and an 18:45 status bar. The accepted prompt says the image is the camera view.
C · sofia_t2Fake status-bar clockA faint 16:45 at the top. Next prompt banned status bars, clocks and UI overlays.
C · s2_t1The suite didn’t readNext prompt filled the frame with the open window and more suite cues behind her.
D · s3_t1Blocking didn’t readBoth twins on one bed. Next prompt asked for two separate single beds, which later broke continuity with s1 and s2.
A · maya_t1Arm in frame, readable signsNext prompt dropped the arm-length wording and asked for blurred, unreadable signs.
A · maya_t3Arm in frame, weaker faceNext prompt named Maya and gave a full face spec with a higher, flattering camera angle.
A · s1_t2Reframe not preferredThe face was pushed higher. Take 1 was restored, byte-identical.
A · s2_t1Identity driftHair pulled back and a different face. Next prompt asked for loose hair exactly as in the reference.
C · sofia_t1Selfie arm across the frameNext prompt kept the arm in the lower-left corner.
Two accepted Balcony images were mirrored locally so Ines faces Clara. That fixed the eyeline without a rejection.
What the numbers say
10 of 12clips said their line on the first take
Both misses were Twins s2. Veo filled the pause between two speakers with words nobody wrote, “um, challenge $100” and then “mm-hmm”, for WER 0.300 and 0.200. A 0.746 s mute took the kept take to 0.000.
36%of paid stills never reached video
10 of 28 stills were rejected before a clip could be spent on them. No clip was re-shot for picture reasons. The one re-shoot, Twins s2, was for audio.
10.22 stof pitch drift inside one character
Sofia’s three clips match each other at only 0.651 to 0.719 cosine, and her s3 sits closer to Clara at 0.774. Maya’s clips hold together at 0.775 to 0.824, above the 0.711 of the nearest other woman. Same model, same prompt style, very different consistency.
23defects the gate never scored
The gate and the reviewer agree on the order, A, D, C, B. But the reviewer logged 23 timestamped problems no gate check measures, like a bed that changes between shots, a line that contradicts itself and a banner that says you while she says me.
0.00points of margin on The Upgrade
C passes at exactly 95.00 with its voice check at 0 of 5. One more lost point anywhere would fail it.
HTTP 402is where the run stopped
Balcony s3 and s4 were refused because the shared balance couldn’t cover them. Auto top-up was off. That left B interim at 81.86 with the brand never spoken. Money stopped it, not quality.
07 / The bill
Every paid call, for members.
The ledger matches the 29 image and 13 video reservation lines treg returned.
This ledger covers metered API calls only. Local compute, agent subscriptions and human time aren’t priced here.
08 / Timeline
One night, then a morning of fixes.
The first scrape ran at 22:56. By 23:37 every clip that would ever exist had been generated. The next morning’s work was all local edits, re-gates and this record. Times are local, UTC+7.
bookingwolf.com scraped (anyapi.web.scrape)
Probe still submitted (Gemini 3 Pro Image)
Probe Veo clip submitted; finished in 106.8 s
Booking Wolf logo saved for the end card
First campaign request (maya master take 1); 4 concepts shot in parallel
First production Veo clip submitted (Twins s1)
Balcony s3 and s4 refused: HTTP 402 insufficient_balance
Twins s2 take 2 delivered (take 1 had WER 0.300); last paid Veo call 23:37:01 (Balcony s2)
Twins s2 'mm-hmm' muted locally
Balcony interim assembled and gated (81.86)
Blind adversarial review written (23 defects, 3 fixes)
First campaign page built (disclosure at y 1540)
brand.json set disclosure null; assemble.py/page.py/page_extra.json updated
All four masters re-assembled without the label (07:04:18-07:04:57)
qa.py Label gate made relative to noise bands; all four re-gated 07:06:21-07:06:55
Campaign page rebuilt
09 / Disclosure and rights
What this page won’t claim.
Five people appear in these ads and none of them exist. Here’s the rest of the fine print, in plain words.
Fictional people
Maya, Clara, Ines, Sofia and the twins are AI-generated fictional characters. Gemini 3 Pro Image made them from text-only prompts and Veo 3.1 Fast animated them. Nobody was filmed, and the voices are Veo’s own generated speech, not clones of anyone.
Dramatised prices
Every price is an example figure from bookingwolf.com’s homepage copy, captured 23 Sep 2026. The site shows its own prices only after a sign-in. So none of the numbers was checked against a booking, and none of them is verified.
No results to report
These ads haven’t been run as paid media. No audience, click or booking data exists for them, and this page doesn’t predict any. The only scores here are an internal QA gate and a blind creative review.
No endorsement
Booking Wolf is named because the ads are about its site. Booking Wolf hasn’t endorsed this page, and none is implied. Its logo appears only inside the ad frames, never as page artwork.
Label outside the frame
The first cuts carried a burned-in AI line where the TikTok and Reels overlay hides it, so it was removed. The disclosure now sits next to every player. Any real upload should also use the platform’s AI-content label.
Included with Lifetime
Get the pipeline kit with Lifetime.
Also included with Lifetime, separate from the core service, is a personal licence to the MakeRoll UGC pipeline kit. After payment the bot sends your member key, and your AI agent or one installer sets the kit up. You run ./ugc run on your own Mac with your own treg.to account, which pays the AI costs. The kit is the animated-story version of the pipeline. The photoreal UGC ads in MakeRoll's case studies came from a separate ad workflow that isn't part of the kit. The setup guide, system requirements and licence are on makeroll.com/ugc-pipeline/. The kit is final once delivered, as the refund policy explains. Setup help from Michael for your own site is an optional extra, on request. Nothing about results is promised, and that includes views, sales and bookings.
One payment for the lifetime plan. No weekly renewal.
24 months
Lifetime means the lifetime of the MakeRoll core service, with at least 24 months.
Members or partners only
This study’s metered spend for four concepts. You pay your own AI generation and tool costs.
Built in public, one brief at a time
MAKE/ROLL.
MakeRoll is a weekly AI video build room. You get a current brief, exact prompts, critique on the cut and a reason to finish one more piece of work each week.
