The reason most people never test their ads properly is not laziness. It is that a variation used to cost a shoot. When making creative number two means another day of filming or another creator fee, of course you run one ad and hope. The moment a variation costs a sentence and a few minutes of render time, testing stops being a luxury and becomes the normal way to work. That shift is what this setup buys you.
Twenty random ads teach you nothing. Twenty ads where nineteen things are held identical and one thing moves teach you exactly which thing moves the numbers. That is the whole discipline, and it is the part the tools do not do for you. Before you generate anything, decide which single variable you are testing this round.
Open claude.ai and start a fresh chat. This first prompt is not the creative, it is the ground truth everything else gets generated from. Spend five minutes here and the twenty scripts come back specific instead of generic.
I am about to generate a batch of short video ad scripts with you, and before we write anything I want you to interview me and then write the brief we will both work from. Ask me these things one message at a time so I can answer properly: what the product is and what it literally does, who buys it and what their day looks like before they find it, the single strongest outcome a customer gets, the price and how it is billed, the two objections that kill most sales, and the platform these ads will run on.
When I have answered, write the brief back to me as a short document with these labelled sections: product, buyer, core outcome, price, objections, platform, and a list of the specific claims I am allowed to make. In the claims section, only include things I actually told you or that appear in material I paste in. Do not add benefits, statistics, guarantees or emotional promises that I did not give you, and if a section is thin, say which question I answered weakly instead of filling the gap yourself.
Keep the brief under 400 words and write it in plain spoken language rather than marketing language, because everything you generate from here has to sound like a person talking, not a landing page.This is the one you came for. It writes twenty scripts from the brief you locked, isolates one variable, and hands you a numbered table so you can tell afterwards which line did the work. Paste it into the same chat.
Using the brief above, write me twenty short video ad scripts for the same offer. The variable I am testing this round is the OPENING HOOK. Every other element must be identical across all twenty: same offer, same core outcome, same single claim in the middle, same closing ask, same length.
Each script is spoken by one person to camera and must run under twenty seconds when read aloud, which is roughly forty to fifty words. Structure every one the same way: one hook line, one line naming the problem in the buyer's own words, one line landing the product as the fix, one closing ask. Write the way people actually talk, including contractions and slightly imperfect sentences, and do not use ad language like "unlock", "elevate", "game-changer" or "revolutionary".
Make the twenty hooks genuinely different in KIND, not twenty rewrites of one idea. Cover at least these angles and label which is which: a direct outcome statement, a specific number, a question, a mistake or warning, a contrarian take, a personal confession, a before and after, a callout to a specific type of person, a curiosity gap, and an objection said out loud first. Two per angle.
Output as a numbered list where each entry shows: the angle label, the hook line on its own, then the full script. After the list, add a plain table with three columns, script number, angle, and hook, so I can paste it into a spreadsheet and record results against it. Do not generate any video yet and do not send anything anywhere until I say go.Read all twenty before you render a single one. Reading takes three minutes. Regenerating twenty videos because the tone was off costs credits you cannot get back. Delete anything that makes a claim your brief did not authorise, then tell Claude which numbers survived.
Arcads is the tool that turns each script into a talking-head video with an AI actor. Connecting it to Claude means you never leave the thread: the scripts you approved go straight through to be rendered. There are two routes, and you only need one.
Add Arcads as a custom connector once and generate from a normal chat. Nothing to install, no terminal. Pick this if you have never opened Claude Code.
Install the official Arcads plugin, which brings the same connection plus ready-made ad skills for competitor research and hook cloning. Pick this if you want the research half too.
https://mcp.arcads.ai. Save it./plugin marketplace add arcads-ai/skills, then /plugin install arcads@arcads, then /mcp to sign in. The plugin ships its own connection, so installing it wires up Arcads for you. The repo is public at arcads-ai/skills.Send scripts 1 to 5 from the approved list to Arcads using the actor and product we already approved, and keep everything except the hook line identical between them so the test stays clean.
Before you send anything, show me in one line what you are about to generate and wait for me to confirm, because every generation spends credits I cannot get back. After they render, give me a plain list mapping each finished video to its script number, its angle label, and its hook line, so I know exactly what I am about to test.
Then stop and wait. Do not render the remaining scripts until I have watched these five and told you to continue.A test only pays off if you write down what happened. Run the batch, give it enough spend and enough time that the numbers mean something, then bring the results back to the same thread and let Claude do the reading.
Here are the results from the batch you wrote, pasted below as rows of script number, spend, impressions, three-second views, clicks, and conversions.
Read them against the angle labels you assigned earlier and tell me which ANGLE won, not only which individual video won, because I care about the pattern I can repeat rather than one lucky script. Group the results by angle, show the average for each group, and say plainly where the sample is too small to trust. If two angles are within noise of each other, say so instead of picking a winner to be helpful.
Then write the next round: fifteen new scripts that all use the winning angle, varying only the specific wording of the hook, plus five that test the runner-up angle so I keep a small bet on the alternative. Keep the same offer, actor, structure and closing ask as before.https://mcp.arcads.ai.