Claude + Higgsfield guide

Check your Reel before you post it

Higgsfield's free Virality Predictor shows where attention sags, second by second. Here is how to run it with Claude Code and turn the graph into exact edits.

The problem

You spend hours on a Reel, post it, and find out in the first hour that people left at second 2. Higgsfield now has a tool that shows you that before you post. It is called the Virality Predictor, and it plugs into Claude Code.

What it does

You give it a clip up to 15 seconds. It comes back with:

a score from 0 to 100
how strong your hook is
a predicted attention curve, second by second
a heatmap of how a brain is predicted to respond

It is a prediction from an AI model, not real viewer data. Treat it like a second pair of eyes, not a verdict. It is also free right now: Higgsfield lists it as an experimental preview that does not use credits. That can change, so check your balance the first time.

The catch

It tells you where attention drops. It does not tell you why, or what to change.

That gap is what the rest of this guide fills.

Step 1 · 5 minutes, once

Connect Higgsfield to Claude Code

Make a free account at higgsfield.ai.
In your terminal run:

claude mcp add --transport http higgsfield https://mcp.higgsfield.ai/mcp

Open Claude Code, type /mcp, pick higgsfield, and sign in.

On claude.ai instead: Settings, Connectors, Add custom connector, name it Higgsfield, paste the same URL. You can also skip Claude and upload the clip on the Higgsfield site. Claude just makes the next steps faster.

Step 2

Run your hook first

Cut the first 15 seconds of your Reel. The hook decides everything, so test it alone before the rest. Upload it, run the predictor, and look at two things:

the dip right after second 1
whether the curve climbs back by second 3

Every clip dips a little at the start. A small dip that recovers fast is fine. A big dip that stays flat means the opening is not grabbing anyone. Open on your most exciting moment instead of building up to it.

Step 3

Find the flat spots

Longer than 15 seconds? Cut it into 15 second pieces and run each one. Start each new piece about a second early so the clip-start dip does not land on a new second. On each graph, mark the stretches of 2 or more seconds where the curve sits clearly below its own peak. Those are your flat spots.

Each piece is scored on its own scale. Compare shapes, not levels across pieces.
A flat spot over your end card or CTA is normal. Ignore it.

Step 4

Turn each flat spot into one edit

For each flat spot, grab the frames from those seconds and ask Claude what is on screen and what to change. Most flat spots come from the same few things:

the same shot held for 3 seconds or more
a text-only card
half the frame empty
your face on screen with nothing else moving
a long sentence with no payoff word on screen

Something new every second or two keeps the curve up. The download below has the exact prompt we use.

The kit · 1 of 3

The Flat-Spot Fix Prompt

Paste your frames and the flat-spot seconds into Claude with this. You get one concrete edit per flat spot.

I ran my Reel through a video attention predictor. Here are the frames from a stretch where the predicted attention goes flat.

Reel topic: [one line]
Flat spot: seconds [start] to [end]
Hook dip: [what you saw after second 1]

Look at each frame. For this flat spot, tell me:
1. What is on screen, in plain words.
2. Why attention probably sags there. Pick from what the frames show: same shot held too long, text-only card, empty half of the frame, face with nothing else moving, repeated layout, long sentence with no payoff word on screen. If nothing looks wrong, say so. Do not invent a reason.
3. ONE concrete edit with the exact second it applies to. Good: "At 23 to 25 s, swap the second talking head for the screenshot you are describing, full screen." Bad: "make it more engaging."

Keep it short. One edit per flat spot.

The kit · 2 of 3

The Hook Checklist

Ten yes or no checks for your first 3 seconds. Fix every no before you post.

01Is there a number or a clear result in the first 5 words?
02Is the hook 18 words or fewer?
03Is there movement on screen at frame 0?
04Is there a 2 to 6 word text headline in the first second?
05Is the first spoken word on screen within 0.25 seconds?
06Does it open on the most exciting moment, not a build up?
07Is the thing you are talking about visible, not just described?
08Does anything new appear by second 2?
09Does your face share the frame with something other than a timer?
10Is the payoff word shown on screen when you say it?

The kit · 3 of 3

The Usual Suspects

Flat spot	Fix
Same shot held 3+ seconds	Cut or punch in 10 to 15 percent
Text-only card	Add the thing the text describes, moving
Half the frame empty	Fill it or crop in
Face with nothing else happening	Go full frame on the raw take, or cut to the proof
Long sentence, no payoff on screen	Show the payoff word the moment you say it

Step 5

Fix, re-run, compare

Fix the flat spots and run it again. Compare where the flat spots are, not the overall score. Different edits of the same Reel often land within a few points of each other, so the score is a bad way to pick a winner. The shape of the curve is what to watch.

What it cannot do

It only sees the video. Not your caption, your audio trend, your thumbnail, or when you post.
It cannot tell you a Reel will or will not go viral.
It is a prediction, not real retention data.

Unlock the Fix Prompt and Hook Checklist

The prompt, the 10-point checklist and the Usual Suspects table. Free, opens instantly.