Claude Skills pipeline turns competitor ad research into ad scripts

Skills to make the mfing bills https://x.com/i/article/2114758150402527232

I Built 5 Claude Skills That Run My Creative Research (With the GetHookd MCP Prompts)

Claude isn't a copywriter. It's a creative team the second you give it real ads to look at.

NURA BEAM's 74-second arm-fat ad had run 133 days from a page with 521 active ads when I pulled it into Claude. One chat turned it into a breakdown, a reusable framework, a rebuild plan and a shoot-ready script for a different product.

If Claude keeps writing you generic ads, it's because you gave it nothing to work with.

So I stopped writing one-off prompts and packaged the work into 5 skills I run in order: research, brand DNA, ad rebuild, image ad, script. Each one below has the job, the mechanism, a real ad, and the full prompt word for word.

The mistake: treating Claude like a vending machine

Most people open a blank chat, type "write me 10 hooks for my sleep supplement," and get the same 10 hooks every other brand got.

The output is only as specific as the context and the ads you feed in.

The fix isn't a better magic prompt. It's a fixed pipeline where every step hands the next one real material: live ads, real customer words, a proven structure.

> Bookmark This - The 5-skill pipeline
1. Research skill: pull what's already spending in your niche
2. Brand DNA skill: build the Beliefs Ladder every ad climbs
3. Ad rebuild skill: borrow a winner's decisions, change its surface
4. Image ad skill: one winning static into five branded variants
5. Script skill: a shoot-ready script with 5 hooks

Setup (once)

- Run everything inside one Claude Project. Every time you finish a file, save it to the Project's knowledge. Every chat you open in that Project after that starts with all 4 files loaded, so you never paste them again

- Connect the GetHookd MCP so Claude can pull the competitor ads that are spending right now

- Pull your own customer data first. Your reviews, your support tickets and your post-purchase survey answers

Skill 1: The research skill

Mechanism: spend is the market's vote. Reviews and Reddit tell you what people say. Ads that have been running for 30+ days tell you what's actually making money right now.

Skip this step and Claude pitches you angles your competitors burned through six months ago.

Example: Holistic Wellness Diary, 220 days live, opening on "Healo is a SCAM. At least, that's what I thought when my daughter ordered it for me without asking."

That's the kind of angle a keyword search for "castor oil pack" won't explain to you. The research skill makes Claude name the angle, the mechanism and where the click goes for every ad.

> Steal This - Research skill prompt
Now use GetHookd to pull the top 30 active ads in my niche. Search for [YOUR PRODUCT TYPE] and for [THE MAIN PROBLEM IT SOLVES], US only, running for 30+ days, sorted by days active, max 2 ads per brand. Transcribe the video ads.

For every ad, tell me the angle, the hook, the mechanism it uses to explain the problem and the fix, the awareness level it's written for, and where the click goes (advertorial, listicle, quiz, VSL or product page).

Then tell me which angles every brand is running, which angles nobody is running, and the one claim that shows up in almost every ad. Add all of it to the Deep Research doc as section 6, What's Spending Right Now, with the GetHookd share link next to every ad.

Skill 2: The brand DNA skill

Mechanism: arguments sell, words don't. Every ad that scales takes the buyer somewhere: from where their head is right now to the belief that makes buying the obvious next move.

The skill builds 4 files once (Deep Research doc, Buyer Persona, Offer Strategy, Beliefs Ladder). The Beliefs Ladder is the one that turns into ads.

Example: High Performance Nutrition, 320 days live, opening on a doctor being asked "how long have you been taking NAD?" One format, one belief: a doctor uses it himself.

> Steal This - Beliefs Ladder prompt
Perfect. Every ad has one job. It takes the beliefs a buyer has today and moves them, one step at a time, to the beliefs that make buying our product the obvious decision.

Read the Deep Research doc, the Buyer Persona and the Offer Strategy. Then write the Beliefs Ladder for [YOUR PRODUCT]: the beliefs a buyer absolutely has to hold before they buy. No more than 6. Write each one as an "I believe that…" statement, in the order they need to be installed, starting with the belief that's furthest from where the buyer is today.

Under each belief, give me:

The proof that installs it, pulled from the Deep Research doc (a quote, a study, a story or a demo)
The objection it kills
The ad format that carries it best (UGC, podcast clip, founder video, listicle, advertorial or static)

Then turn every belief into one ad concept with a hook. Six beliefs, six concepts.

Every rung on the Beliefs Ladder is a different reason to buy, so every rung is a different ad. Meta's delivery in 2026 rewards ads that are actually different from each other, and 40 versions of the same ad with new hooks don't count.

> Save This - Start every new chat in the Project like this:
Read the Deep Research doc, Buyer Persona, Offer Strategy and Beliefs Ladder before you answer. Then [WHAT YOU NEED].

Skill 3: The ad rebuild skill

This is where the article gets specific. If you only want the big idea, it's above. From here on it's the exact prompts I run.

Mechanism: borrow the decisions, change the surface. Winning ads work because of the decisions underneath (structure, tension, proof type, format), not the words on top.

Example: the NURA BEAM ad from the top. Written for problem-aware buyers in a stage 4 market, it opens on a comment reply, does the effort math, drops a named mechanism in the middle and puts the guarantee before the offer.

The beat structure Claude pulled out of it, as a template:

- 0–3s, Comment reply + problem shown: Bubble: "Does [product] work for [problem]?" Creator shows the problem area

- 0–12s, Pattern-interrupt line + early proof: "[Stranger] saw my [product] and asked [funny misread]. Only against [problem]." Cut to a 2nd creator with "I'm [X] down"

- 12–19s, Credibility claim + product reveal: "This little thing [does what] using [N] [technologies]." Device switched on in hand

- 19–24s, Effort math: "[X] minutes a day equals [big number] [effort they hate]."

- 24–29s, Result + time frame: "I [measurable result] in [days]."

- 29–33s, Future pace: "For the first time I'm excited to [thing they avoided]."

- 33–58s, Named mechanism + visual cutaway: "This is called [mechanism]." Expert setting, animation of how it works

- 58–63s, Visible result: "Look at that difference." Before/after

- 63–67s, Risk reversal: "No [result] in [days]? You get your money back."

- 67–74s, Offer + CTA: "Tired of [problem]? Get [product] at [offer]." Offer card

And what the original could do better, which your version fixes:

1. One offer, not three. The voiceover says 30% off and 2 free gifts, the end card says free upgrade and 2 free gifts, and a Black Friday badge has been on the ad since June.

1. Show the product sooner. The first line takes 9 seconds and the device isn't confirmed on screen until 18 seconds. Get the line under 4 seconds and the product into the first 3.

1. Prove the claim on camera. Put a tape measure on screen for the inch, give the expert a name and a title, and cut "clinically proven" and "33%" unless you can back them up. Claims like these get health ads rejected.

> Steal This - Ad Analysis + Rebuild Plan prompt
BEFORE YOU START
If I gave you a GetHookd ad ID instead of a breakdown, run the Step 6 gather steps yourself first (get_ad, then get_ad_media for the hook and runtime stills).

Then check whether the structure works beyond one brand: call list_similar_ads(ad_id, per_page=6). It matches on ad copy, not structure, so call get_ad on the 3 closest results and read their transcripts. Count how many DIFFERENT landing domains run the same beat structure. Pages that send traffic to the same domain are one company. Tell me the count and the ad IDs.

AD ANALYSIS + REBUILD PLAN PROMPT

You are a senior DTC creative strategist and consumer psychologist. Below is a neutral, descriptive breakdown of a competitor ad. Your job is to work out why it works, extract its structure as a reusable framework, and plan how to rebuild that structure for my product using my Brand DNA Files.

The core principle: winning ads work because of the decisions underneath, not the surface. We borrow the decisions (structure, tension, proof type, format choice). We change the surface (words, visuals, claims, people, examples, product details). And we fix what the original does badly, so our version beats it.

AD BREAKDOWN
[PASTE THE FULL BREAKDOWN FROM STEP 6 HERE]

RULES
- Interpret, don't describe. Never repeat the breakdown back; explain why each element works.
- Anchor every point to a timestamp from the breakdown.
- Do not invent anything that isn't in the breakdown. If an element is missing (no mechanism, no offer, no objection handling), write "Not present" and note whether its absence is a deliberate choice or a weakness.
- In Part 3, use only facts, proof and customer language from my Brand DNA Files. If something we need doesn't exist there, flag it as a gap. Never fill a gap with invented proof or claims.

OUTPUT (in this exact order)

PART 1: WHY THE AD WORKS

Awareness: which awareness level the ad is written for, and what it assumes the viewer already knows or believes.
Market sophistication: which stage (1 to 5) the ad is written for, and how you can tell (new claim, bigger claim, mechanism, new mechanism, identity).
The single promise: the one thing the ad promises, in one sentence. If it needs an "and", identify which promise leads.
The hook: what stops the scroll in the first 2 seconds, visually and verbally, and what tension or curiosity it opens.
Problem & agitation: the problem the viewer is made to feel, the emotions tapped (fear, guilt, frustration, embarrassment, desire), and any failed alternatives shown.
Mechanism: the problem mechanism (why the problem exists) and solution mechanism (why this product works differently). How believable it is made.
Objections: the first real objection the ad addresses (exact words + timestamp) and how it's answered, then any others.
Proof: each piece of proof, where it appears, and what claim it actually supports.
Belief journey: the beats in order with timestamps, and which belief each beat installs on the way to the CTA.
Format: why this format (UGC, demo, founder, interview, etc.) makes the promise easier to believe.
Offer & CTA: the reason to act now, the exact CTA and, if a landing page is known, whether it continues the same promise.
Retention: how pacing, editing, text and visuals keep people watching.
Funnel stage: top of funnel (cold traffic, sells the problem and the mechanism) or bottom of funnel (sells the offer), and why.
Top 3 performance drivers: the three decisions most responsible for this ad working, ranked. These are what we must not lose in the rebuild.
What it could do better: 2 to 3 concrete weaknesses (mixed offers, slow product reveal, weak proof, risky claims) and how our version fixes each one.

PART 2: REUSABLE FRAMEWORK
Rewrite the ad as a universal template any brand could reuse. Replace every brand-specific element with a placeholder like [product], [problem], [failed alternative], [mechanism], [proof], [outcome], [customer type], [offer].

PART 3: REBUILD PLAN FOR MY PRODUCT

Fit check: does this structure suit my product? Compare the ad's awareness level, sophistication stage, promise type and proof needs with what my Brand DNA Files support. If there's a serious mismatch, say so plainly and explain what would need to change, before continuing.

Borrow vs. change:
| Borrow (structure and decisions) | Change (surface and execution) |

Gaps: anything the source ad relies on that we don't have (proof, offer, mechanism, results), with the closest real alternative from my Brand DNA Files.

When you're done, stop. Do not write the script yet. I will review the rebuild plan first.

Structures that run on 2 or more different landing domains are the safest ones to reuse. For NURA BEAM, list_similar_ads found 3 more pages with the same structure, all sending traffic to the same site, so it's one company: a strong test, not a sure thing.

Want these skills on tap?

All 5 skills read live ads, transcripts and share links straight from GetHookd inside Claude. If you run Meta ads, the GetHookd MCP is what makes Skills 1, 3 and 4 work without a single screenshot.

Skill 4: The image ad skill

Mechanism: proven concept in, on-brand creative out. The bottleneck isn't "can the machine make a beautiful image?" anymore. It's "which image is WORTH making?"

Example: Deborah Rustad, 204 days live. A plain photo of an unmade bed in a mountain room, headline "I was shocked by the findings," copy opening on "I just spent $6,800 on a sleep clinic in the Swiss Alps."

> Steal This - The Static Recreator (one winner → five branded variants)
Take this single winning static: [PASTE GETHOOKD SHARE LINK or AD ID]. Pull it with get_ad on the GetHookd MCP. Then: (1) full teardown — concept, layout grid, hook text verbatim, color psychology, emotional trigger, and the ONE thing that makes it work; (2) write FIVE complete image-generation prompts recreating this concept for my brand [PASTE BRAND INPUTS], each riding a different angle: pain-led, aspiration-led, social-proof-led, mechanism-led, and offer-led; (3) for each, the overlay copy in my customer's language + placement; (4) tell me which variation to test first and what result would prove the concept transfers to my market. Same rules as always: product accuracy via real-photo compositing, realism over stock vibes, 1:1 default.

Generate the visual clean and add the overlay text in Canva. Image models still fumble exact copy, and your headline is too important to gamble on.

Skill 5: The script skill

Mechanism: customer words beat copywriter words. The script skill writes only from the rebuild plan and the Brand DNA Files, so every claim traces back to proof.

Stay in the same thread as Skill 3 and run it.

> Steal This - Ad Script prompt
AD SCRIPT PROMPT

You are a world-class DTC creative strategist and direct response scriptwriter. Using the approved rebuild plan above, the competitor ad breakdown and analysis in this chat, and my Brand DNA Files, write a shoot-ready paid social ad script for my product.

BEFORE YOU WRITE
If anything essential is missing or unclear from the rebuild plan or the Brand DNA Files, ask me your clarifying questions in one message and wait for my answers. If nothing is missing, start writing.

RULES
- Follow the beat structure from the approved rebuild plan exactly, in that order. If I edited the plan, my edits win.
- Fix every weakness of the original listed in the plan.
- Write in the language of the market stated in the Brand DNA Files.
- Match the competitor ad's runtime within roughly 10%. Pace spoken lines at about 2.5 to 3 words per second.
- Borrow structure, never surface. Do not reuse or lightly rephrase any line, phrase or visual idea from the competitor ad.
- Use only facts, claims and proof from the Brand DNA Files. No invented results, numbers, studies or testimonials.
- Use real customer language from the Deep Research doc wherever it fits. Customer words beat copywriter words.
- Write for the ear: short sentences, natural rhythm, words a real person would say on camera. Read every line as if spoken aloud.
- No placeholders. Every line is final copy.

OUTPUT (in this exact order)

PART 2: HOOK VARIATIONS
Write 5 alternative hooks for the first 2 to 3 seconds. Each must flow naturally into the same body script. Use a different hook type for each (e.g. problem callout, bold claim, curiosity, pattern interrupt, customer quote). For each hook give:
- Hook type
- Spoken line
- On-screen text
- Opening visual

PART 3: CREATOR & EDITOR NOTES
- Delivery: tone, energy, pace, what to emphasize
- Setting, wardrobe and props
- Shot list: every shot and b-roll clip needed, so it can be filmed in one session
- Editing notes: caption style, cuts, text animation, music direction

You'll get the full script in a table, 5 alternative hooks that flow into the same body, and notes for your creator and editor.

> Steal This - Hook check: are your hooks what everyone else is already saying?
Take the ads from my Step 1 list and the competitor ads from my Competitor Breakdown. For any ad without a transcript, call transcribe_ads (free) and wait 60 seconds. Then call get_ad_hooks on each ad to get its spoken opening line. Put those hooks next to my 5 new hooks and flag any of mine that say the same thing or follow the same pattern.

Then call get_creative_breakdown(shop_id=...) for my top 2 competitors and show me their spoken_hook mix, with how many ads were analyzed.

If your hook is what 10 brands already say, rewrite it. If most of your competitors open with a question, open with something else.

The short version

Research → brand DNA → rebuild → image → script. Each skill hands the next one real material, which is why the output stops sounding like everyone else's.

If you run Meta ads and want these 5 skills running on real ad data inside Claude, connect the GetHookd MCP: gethookd.ai/mcp

API & MCP access is included with every annual plan and with the Team/Agency monthly plans.

Save this.