Replace Your Ad Agency With Claude Cowork

Bottom line

Brief one Claude cowork session and walk away with ten finished video ad cuts, engineered to persuade and ready to test and scale. No agency, no retainer, no shoot day.

This guide is reviewed for clarity, service accuracy, and AI-search readability. The next quarterly content review is tracked internally before unsupported metrics or client proof are added.

You're not buying an ad. You're buying a system that prints ads.

A traditional 30-second commercial costs $10,000 to $50,000 to produce. AI tools now make comparable creative for $100 to $1,000. That's a 95%+ drop in cost. It's not a small saving. It changes what's possible.

When one ad costs minutes instead of money, you stop betting everything on a single idea. You ship a batch, let the platform's algorithm vote, and pour budget into the cuts that already work. That's the whole game now. Creative quality drives 47–56% of advertising effectiveness, more than targeting or media optimization. So the team that tests the most good creative wins.

This is the exact system we run at Hypd to do that with one Claude "cowork" session. You brief it once. It acts as your creative director, strategist, and copywriter. It directs the shots and hands you ready-to-render prompts. You walk away the same day with a batch of finished video ad cuts, built on the same buying triggers a senior strategist would reach for.

No agency. No retainer. No shoot day.

What you'll walk away with

A repeatable cowork session that turns one brief into ~10 distinct ad angles
Copy-paste prompts for every stage: brief, direction, script/hooks, and render handoff
Two blueprints that cover both halves of your traffic: cold strangers and warm prospects
A persuasion checklist baked into every cut
A simple test-and-scale rule so the data, not your ego, picks the winners

The tool stack and account setup

You need three accounts. That's it.

1. Claude (the cowork brain). This is your agency in one chat. It runs the brief, the creative direction, the scripts, the hooks, the storyboard, and it writes the render prompts you feed the video engine. You never need to know how to "prompt a video model." Claude writes those prompts for you.

2. An AI video render engine. This generates the actual footage. Pick one to start:

Veo 3.1 — audio-native cinematic generation. It makes the picture and the synced sound in one pass, so you often skip a separate voiceover step.
Seedance 2.0 — unified audio-video joint generation, strong value at scale when you're rendering ten cuts at once.
Kling 3.0 — multi-shot storyboard support and native 4K, good when your ad is several shots stitched together.
Sora 2 — physics-first, the most realistic motion when products move, pour, or get handled.
Runway Gen-4.5 — the most creative control if you like to direct shot by shot.

For ad creative, start with Veo 3.1 or Seedance 2.0 because the native audio means fewer moving parts.

3. An ad account on Meta, TikTok, or YouTube, with your pixel/tracking installed so you can read results.

Optional but useful:

A voiceover tool. If Veo/Seedance native audio covers your line, you don't need this. For a scripted, branded VO read, a text-to-speech generator (such as ElevenLabs) gives you a clean take.
A lightweight editor (CapCut works free) to trim, stitch shots, and burn in captions. Caption it for sound-off viewing.
A creative testing layer. Meta and TikTok Ads Manager are enough to start. Tools like Motion, Madgicx, or Marpipe help once you're testing dozens of cuts a month and want to know why a winner won.

One rule on quality: research from NielsenIQ found that ads that look obviously AI-generated trigger negative brand perception. So use AI for production speed, but keep the creative native and authentic. UGC-style footage beats glossy studio polish on nearly every metric. Tell the engine to make it look like a person filmed it.

The cowork kickoff: the command and the brief

Open a new Claude session. Paste this. Fill in the brackets. This is what launches the whole thing.

prompt
You are my in-house creative director and direct-response ad strategist for this session.

We are running a "cowork" sprint. By the end, I want a batch of finished, post-ready video ad cuts I can launch today.

Here is the brief:
- PRODUCT: [what you sell, in one line]
- OFFER: [price, discount, guarantee, bonus, deadline]
- CORE ANGLE: [the one big idea — e.g., "soda that's actually good for your gut"]
- AUDIENCE: [who they are, the exact pain they feel, where they scroll]
- AWARENESS: [cold = never heard of us / warm = on our list or clicked before]
- FORMAT: [UGC selfie / product demo / founder-to-camera / cinematic]
- ASPECT RATIO: [9:16 vertical / 1:1 square / 16:9]
- DURATION: [e.g., 15s]
- PLATFORM: [Meta / TikTok / YouTube]
- BRAND RULES: [tone, colors, must-say lines, must-not-say words]

Confirm you have the brief. Ask me ONLY the questions you genuinely need before we move to creative direction. Do not write scripts yet.

The brief is the whole job. Garbage in, generic ads out. Be specific about the pain and the offer. Everything downstream inherits this.

The stage-by-stage workflow

Five stages, all in one session, all the same day:

Brief — you load the kickoff command above.
Creative direction — Claude proposes distinct big ideas (territories), not variations of one.
Script and storyboard — you pick territories; Claude writes the second-by-second script, alternate hooks, and a shot list.
Render handoff — Claude converts the shot list into paste-ready prompts for your video engine. You render.
Review — you watch the finished cuts, caption them, and launch the batch.

Here are the copy-paste prompts for stages 2 through 4.

Stage 2 — the directing prompt

prompt
Now act as the creative director. From the brief, give me:

1. Three distinct creative TERRITORIES (genuinely different big ideas, not three versions of one).
2. For each territory: the buyer's awareness stage, the ONE emotion we're pulling, and the SINGLE objection we're killing.
3. The visual world for each (setting, talent, pacing, mood) in 2-3 lines.

Keep it tight. I'll tell you which territories to develop into full cuts.

Stage 3 — the script and hook prompt

prompt
Take territory [number/name]. Write it as a [DURATION] [ASPECT RATIO] [FORMAT] ad for [PLATFORM].

Use this direct-response skeleton, second by second:
- 0–3s HOOK: stop the scroll. Visual + spoken line. The first frame must earn the second.
- 3–7s PROBLEM / AGITATION: name the pain in the buyer's own words.
- 7–11s PROOF / MECHANISM: why it works — demo, result, or reason to believe.
- 11–14s OBJECTION KILL: remove the #1 reason they'd scroll past.
- 14–15s CTA: one clear next step tied to the offer.

Give me:
A) the full script with on-screen text AND spoken voiceover,
B) 3 alternate HOOKS (the first 3 seconds only) for the same script,
C) a shot list / storyboard, one line per shot, with camera move + action.

The first three seconds carry the ad. Meta's first-3-seconds hook rule is non-negotiable, which is why you generate three hook options for every single script and test them against each other.

Stage 4 — the render handoff prompt

This is the trick most people miss. You don't write video prompts. Claude does.

prompt
Turn the shot list above into ready-to-paste prompts for [Seedance 2.0 / Veo 3.1 / Kling 3.0].

For EACH shot, output one prompt block containing: subject + action, camera move, lens/framing, lighting, mood, setting, and any spoken line for native audio.

Hard constraints:
- [ASPECT RATIO], [DURATION] per shot
- photoreal UGC look (looks like a person filmed it, NOT glossy or obviously AI)
- product shown clearly and accurately as: [precise product description]

Output each shot's prompt in its own numbered code block so I can paste them straight into the engine.

Paste each block into your render engine. Stitch the shots. Caption it. Done — finished cut, same day.

The persuasion-engineering layer

A good-looking video isn't an ad. An ad does persuasion work. Every cut you ship should run on the same direct-response levers a senior strategist would reach for. The skeleton in the script prompt above bakes them in, but check every cut against this before it goes live.

The senior-strategist checklist:

Two blueprints: cold vs warm

The same system has two jobs. A cold prospect has never heard of you. A warm one already clicked once and stalled. You can't sell them the same way. Cold prospecting audiences are mostly Unaware and Problem Aware. Your retargeting pool is stacked with Product Aware and Most Aware people. The ad has to meet each where they are.

The only thing that changes is what you paste into Stage 3. Add one of these blocks.

COLD traffic — scroll-stop and earn the first click:

prompt
AUDIENCE AWARENESS: COLD (Unaware / Problem Aware — they've never heard of us).
JOB: stop the scroll and earn the first click. Do NOT lead with the brand or the offer.

- HOOK: lead with the PROBLEM or a pattern interrupt, not the product.
- Spend the first half on the pain. Introduce the product as the discovery, not the pitch.
- CTA: low-commitment ("see how it works", "learn more"), never "buy now".
- Tone: native, like a friend filmed it on their phone.

WARM traffic — answer the last objection and close:

prompt
AUDIENCE AWARENESS: WARM (Product Aware / Most Aware — they know us, clicked once, didn't buy).
JOB: answer the last objection and close. You CAN name the brand and offer immediately.

- HOOK: call out the hesitation ("Still on the fence?") or open with a fresh proof point.
- Spend the middle on the objection: price, risk, "will it work for me".
- CTA: direct and urgent, tied to the offer ("20% off ends Sunday", "free shipping today").
- Lean hard on proof: reviews, results, the guarantee.

Run both. Same system, two jobs, both handled.

Ship ten, not one

When each ad costs minutes, testing one idea is malpractice. You generate a batch from a single brief and let the market pick.

The batch protocol:

Ask Claude for 5 territories instead of 3 (just change the number in the directing prompt).
For each territory, take the script and its 3 hook variations. That's already 15 openings from 5 ideas.
Render the strongest ~10 cuts across those territories and hooks. Mix cold and warm.
Launch them together so the algorithm can compare them on the same playing field.

Then label everything. You cannot scale what you can't tell apart. Use one naming convention across files and ad names:

prompt
[product]_[awareness]_[territory]_[hook]_[aspect]_v[number]

Examples:
olipop_cold_gut-health_doctor-hook_9x16_v01
olipop_cold_gut-health_soda-swap-hook_9x16_v02
olipop_warm_taste-objection_review-hook_9x16_v03

Now when a winner appears in the data, you know exactly which angle and which hook won, and you can make more like it.

The test-and-scale rule

The winners show themselves in the numbers. Your only job is to read them honestly and act.

The metrics that decide a winner (in order):

Hook rate — 3-second views ÷ impressions. If they don't stop, nothing else matters. Kill weak hooks first.
Hold rate — how far they watch. A strong hook with a cliff-drop means the body is broken.
CTR — are they clicking through.
CPA / ROAS — the only metric that pays the bills. Cost per acquisition, or return on ad spend.

A simple rule that works:

Give each cut a fair shot. Let it spend roughly 1.5–2× your target CPA before you judge it. Less than that and you're reading noise.
Retire losers fast. If a cut burns ~2× your target CPA with no sale, or its hook rate sits at the bottom of the batch, cut it. No ego, no second-guessing.
Scale only proven cuts. Take the few beating your target CPA/ROAS and increase their budget ~20–30% every 2–3 days, or duplicate them into a dedicated scaling campaign. Pour budget into what's already pulling.
Refill the pipeline. Feed your winning angle back into a new cowork session and ask Claude for ten fresh variations of that idea.

This is exactly why the math matters. Kalshi ran a $2,000 AI-generated spot that reached over 20 million impressions. The point isn't the one ad. It's that cheap, fast creative lets you find the breakout instead of guessing at it.

Worked example: an OLIPOP-style prebiotic soda

Here's the full path, brief to finished cut, modeled on the kind of spot you'd open with.

The brief loaded into the kickoff command:

prompt
- PRODUCT: A prebiotic soda that tastes like classic pop but supports gut health (low sugar).
- OFFER: First case 20% off + free shipping, money-back guarantee.
- CORE ANGLE: The soda you're allowed to love — the taste of childhood pop, none of the sugar guilt.
- AUDIENCE: 25–40, health-curious, misses real soda, scrolls Instagram/TikTok at night.
- AWARENESS: COLD.
- FORMAT: UGC selfie, handheld.
- ASPECT RATIO: 9:16.
- DURATION: 15s.
- PLATFORM: Meta.
- BRAND RULES: Warm, playful, real. Show the can clearly. Never claim it cures anything.

Three hook variations Claude returns (first 3 seconds only):

prompt
HOOK A — "I quit soda for two years. Then I found one I'm allowed to drink."
   [Visual: person cracks the can open, close on the fizz]

HOOK B — "POV: it's 9pm and you want a soda that won't wreck your gut."
   [Visual: fridge door opens, hand reaches past energy drinks to the can]

HOOK C — "Dietitians keep telling me to drink this instead of regular soda."
   [Visual: person mid-sip, eyebrows up, surprised it tastes good]

The finished 15-second cut (Hook A, cold blueprint):

0–3s (hook): Handheld, person cracks the can, "I quit soda for two years. Then I found one I'm allowed to drink." Native, phone-shot look.
3–7s (problem): "Regular soda is basically a sugar bomb. But sparkling water is sad and I missed the real thing."
7–11s (proof/mechanism): Pour over ice, close-up fizz. "This one's a prebiotic soda. Tastes like proper pop, low sugar, actually good for your gut."
11–14s (objection kill): Sips, genuine reaction. "And no, it doesn't taste like health food. It tastes like the soda I grew up on."
14–15s (CTA): Can held to camera with on-screen text. "First case is 20% off. Tap to try it."

Claude wrote the script, the three hooks, and the shot list. The Stage 4 prompt turned each shot into a render block for Veo 3.1. You stitched four shots, burned in captions, and had the cut posted that afternoon. Then you did it nine more times.

One-page quickstart checklist

prompt
SETUP (once)
[ ] Claude account open
[ ] Video engine chosen (Veo 3.1 or Seedance 2.0 to start)
[ ] Ad account + pixel/tracking live

EACH COWORK SESSION
[ ] Paste the kickoff command, fill the brief (be specific on pain + offer)
[ ] Answer Claude's questions
[ ] Run the directing prompt → ask for 5 territories
[ ] Pick territories → run the script/hook prompt (get 3 hooks each)
[ ] Add the COLD or WARM block per cut
[ ] Run the render handoff prompt → paste blocks into the engine
[ ] Render, stitch, caption for sound-off
[ ] Run the persuasion checklist on every cut
[ ] Name every file: product_awareness_territory_hook_aspect_vNN

LAUNCH & SCALE
[ ] Launch ~10 cuts together (mix cold + warm)
[ ] Watch hook rate → hold rate → CTR → CPA/ROAS
[ ] Kill anything past ~2x target CPA with no sale
[ ] Scale winners +20–30% every 2–3 days
[ ] Feed the winning angle back into a new session for more variations

A quick note on staying compliant

AI disclosure rules expanded in 2025–2026. New York requires conspicuous AI disclosure in some ads, California's AB 2655 covers political content, and the EU AI Act (Article 52) requires disclosure of synthetic media. Check the rules for your platform and region before you launch. And remember the NielsenIQ finding: keep it native and authentic. The goal is creative that uses AI, not creative that screams it.

FAQ

Do I need to know how to prompt a video model? No. That's the point of the cowork session. Claude writes the render prompts for you in Stage 4. Your job is the brief and the judgment calls, not the technical prompting. You paste what Claude gives you into the engine.

How many ads can I realistically make in an afternoon? From one brief you can have Claude generate 5 territories with 3 hooks each, then render the strongest ~10 cuts. The bottleneck is render time on your video engine, not Claude. Most solo founders ship a full batch of 10 in a single session.

What if none of my first batch wins? That's information, not failure. Look at hook rate first — if nothing stopped the scroll, your hooks or your core angle are off, so go back to the brief and sharpen the pain. If hooks landed but nobody converted, the body or offer is the problem. Run a fresh cowork session targeting the specific weak point. Cheap, fast creative means a losing batch costs you an afternoon, not a budget.

Want the whole system done for you, or want to scale this past your first ten cuts? That's what we do at Hypd.

The Claude Cowork Ad System: Replace Your Agency in an Afternoon