How to Prompt Higgsfield to Fix a Real Ad

Step 1 — Anchor on a frame from your real footage

Don't generate from imagination — the clip will never match your ad. Instead, have Claude pull a still from the video you already scored and use it as the start image for image-to-video generation. Pick the frame that shows the world your new shot must live in. In our session that was the glowing night-drive dashboard at 6.5 seconds:

Extract the best-looking night-drive frame from my scored ad
and use it as the reference image for the new opening shot.

Because the generation starts from your pixels, it inherits your lighting, interior, color grade — in our run it even reproduced the original's watermark style. That continuity matters more than clip quality: the attention model punishes footage that doesn't belong.

Step 2 — Write the prompt: one beat, motion first

Here is the exact prompt from the session:

POV vertical shot inside a dark car at night. The dashboard is
unlit for a brief instant, then an LED ambient light strip along
the dash and doors suddenly snaps on with a vivid blue-purple
glow, light sweeping rapidly across the dashboard in a satisfying
color burst. Camera static, realistic phone-shot look.

Its anatomy, part by part:

The generated clip: the dark beat (left) cost us; the snap payoff (right) matched the real footage almost perfectly.

Step 3 — Let Claude handle models (and plan walls)

You don't need to know the model zoo. In our session Claude asked for kling3_0_turbo, hit a "basic plan required" wall, tried seedance_2_0 — same wall — then fell back to happy_horse_video, which ran on the free tier. Before paying for an upgrade, just say:

That model needs a paid plan — try whatever image-to-video
model works on my current plan before I upgrade.

The budget model was plenty: 3 seconds, 720p, 9:16, generated in about two minutes — and the spliced result posted the best back-half attention numbers of the entire experiment.

Step 4 — Check the clip before you splice it

Have Claude look at the generated clip frame by frame first — it's free and takes seconds. Ours revealed the snap happened 1.2s in, with dead darkness before it. That told us exactly what to trim before grafting.

Look at the generated clip frame by frame. When exactly does
the snap happen, and is there any dead time before it?

Step 5 — Splice it in (Claude does the ffmpeg)

Claude matched resolution, frame rate, and audio, then concatenated the AI hook onto the untouched original — never onto an already-edited cut. Keep your original pristine and let every experiment branch from it:

Trim the AI clip to just the snap moment, graft it onto the
front of the untouched original, and score the new version.

Step 6 — Re-score, read the seam, iterate

The hybrid scored 42 — and the curve pointed at the seam, not the AI footage: the dark beat opened at 0 attention while the back half hit its best numbers of the run (purchase signal 87). We cut the dark beat and re-scored: 22, because the now-instant jump from glowing dash to dark prep footage read as whiplash. Two iterations, two precise diagnoses, each costing one credit instead of a live campaign.

The iteration rules that fell out of the data: fix the seam, not the clip — the transition is almost always what the model punishes. Never remove the context that sets up your payoff. And when a version loses, believe the number: in our session the untouched original beat every hybrid, and that was the real finding. The loop's job isn't to prove your edit was right — it's to stop you shipping the wrong cut.

Three prompt templates to steal

Swap the bracketed parts. Each is one beat, motion-first, style-matched:

POV vertical shot, [your setting from the reference frame].
[Your product] is mid-[transformation] — [the effect] already
sweeping across [the surface]. Camera static, realistic
phone-shot look.

Vertical UGC-style shot, [setting]. A hand enters frame and
[single satisfying action: clicks/peels/snaps] the [product],
which instantly [visible payoff]. Handheld feel, natural
lighting, no cinematic grading.

Start from the reference image. In the first half-second,
[the hero element] [bursts/lights/transforms], then the camera
holds as [the payoff state] settles. No dead frames before the
action.