AI UGC Hook Test

The three versions

Each version is a complete ad. Watch the first three seconds of each one; everything after the cut is identical.

H01 · Talking head14.8 s

The only cup I leave the house with
She opens with the claim, straight to camera.

H02 · POV proof15.1 s

Made at 7:00. It's 11:40.
Lid off at 11:40 and the coffee still steams. The voice says it was made at seven.

H03 · Silent text + ASMR15.1 s

POV: your 11:40 coffee still steams
No voice. Just the lid sound, the steam and one line of text, made for sound-off feeds.

The creator, voice and brand are AI-generated for this sample and labelled as such. This is a spec ad; it has not run with paid spend.

The Claude Code workflow

The whole concept lives in a single brief file. One command turns it into named, ready-to-upload versions, and adding hook number four means adding one line.

$ python3 build/render.py
KEEL_C01_H01_v01_916.mp4  14.80 s
KEEL_C01_H02_v01_916.mp4  15.06 s
KEEL_C01_H03_v01_916.mp4  15.06 s
manifest.csv written · qa.json: all checks passed

brief.json
{
  "concept": "C01_morning-luxury",
  "version": "v01",
  "angle": "Want, not need: the small daily luxury of coffee that still tastes and feels right hours later",
  "hooks": [
    {
      "id": "H01",
      "type": "talking-head",
      "spoken": "Okay, this is the only cup I leave the house with now.",
      "overlay": "The only cup I leave the house with",
      "source": "main"
    },
    {
      "id": "H02",
      "type": "pov-proof",
      "spoken": "I made this coffee at seven. Look at it now.",
      "overlay": "Made at 7:00. It's 11:40.",
      "source": "pov"
    },
    {
      "id": "H03",
      "type": "silent-text-asmr",
      "spoken": "",
      "overlay": "POV: your 11:40 coffee still steams",
      "source": "pov_alt"
    }
  ]
}

1 · Brief

Angle, script, hooks, overlays and CTA live in a single file. Claude Code writes hook variants from the angle; I pick and tighten them.

2 · Creator and voice

The AI creator still and product shots are generated first. The voice is recorded as one take, and the word timestamps from that take drive the captions.

3 · Lip-sync shot

The creator is animated speaking the take, holding the product. Face, mouth and hands are checked frame by frame, and weak takes are redone.

4 · Versions

The script renders every hook with the same body, captions and CTA. Names follow brand_concept_hook_version_format, so the results in Ads Manager map straight back to the brief.

5 · QA + manifest

Resolution, audio, loudness and length are checked automatically, then frames are checked by eye. The manifest lists every file with its hook and a checksum.

File	Hook	Type	Length	Checksum
`KEEL_C01_H01_v01_916.mp4`	H01	talking-head	14.80 s	`4121dddabc`
`KEEL_C01_H02_v01_916.mp4`	H02	pov-proof	15.06 s	`47494ae7e0`
`KEEL_C01_H03_v01_916.mp4`	H03	silent-text-asmr	15.06 s	`04fd5802cb`

Checks on this sample

Picture

All 3 versions are 1080 × 1920, 30 fps, H.264, 14.9–15.1 s, with no scene breaks inside the talking take.

Sound

Every version has a stereo AAC track normalised to -15.0 to -14.8 LUFS (Meta / Reels range), with the music bed kept well under the voice.

Speech

Each export is transcribed by machine and compared against the script. The words match, and a repeated phrase in the raw take was cut out before the edit.

Hands and face

Frames checked at full size: five fingers on every visible hand, a stable grip on the product, the same face throughout and no warping around the mouth.

One AI creator, one body,
three hooks to test.