Score Your Ad with Claude, Step by Step

Step 1 — Connect the MCP (once)

In Claude, open Settings → Connectors → Add custom connector. Name it PreTestAds, paste the server URL, hit Connect, and sign in with your PreTestAds account. That's the entire setup — you never do this again.

https://pretestads.com/mcp

Step 2 — Ask, then drop your video in

Open a chat and say what you want. In the real session it was five words:

i want to score an ad

An upload box appears inside the conversation. Drag your video in (MP4, MOV, or AVI up to 500MB) — scoring starts the moment the upload lands, and Claude polls until the report is ready.

Step 3 — Read the score (and the curve)

About three minutes later, Claude reads back the full report. The test ad scored 47/100 — moderate: a strong 4-second peak and steady payoff, dragged down by a hook that hit literal zero at second 1.

The attention curve, animated: slow build, huge install peak, long glowing payoff.

Step 4 — Let Claude watch the ad

This is the step no dashboard can do. Ask Claude to look at the footage and it calls get_ad_frames — one still per second, free — and maps what's on screen to the exact seconds where attention moves.

Watch it frame by frame and tell me exactly which second to fix.

Step 5 — Iterate: re-cut, or generate new footage

Claude can re-cut the file itself (it trimmed, re-ordered, and spliced with ffmpeg in the real session) — and with a generation MCP connected, it can create footage too. Here the user typed:

use the higgsfield mcp to get the footage

Claude pulled a reference frame from the scored ad, had the Higgsfield MCP generate a matching "lights snap on" opening, spliced it onto the original, and sent the hybrid back for scoring. Each round costs one credit — the fast feedback is the whole point. The full prompting workflow, with templates, is in the Higgsfield iteration tutorial.

Step 6 — Let the number pick the winner

Six versions later, the scoreboard settled it: the untouched original beat every edit — including ones with objectively better hooks. Claude called it plainly and recommended shipping the original. The score is the referee, not the AI's confidence and not your gut.

Why did every edit lose? The full data story — curves for every version, what improved, what broke — is in the case study.