how i cook 30 winning AI ads in 30 mins w/ claude x higgsfield MCP

i timed the last production session.

from opening claude to having 30 finished ad assets ready for upload to meta: 27 minutes.

not 27 minutes of setup and then hours of generation. 27 minutes total. brief, scripts, character generation, product b-roll, animation, static extraction, copy variations. the full creative matrix for a week of meta testing, done before my second coffee got cold.

the brand i ran this for had been paying a creator agency $4,800 a month for 12 assets with a three-week turnaround. twelve. in three weeks. i produced 30 in less time than it takes to watch an episode of something.

this is the exact workflow, step by step, prompt by prompt.

minute 0 to 5: load and brief

open claude. load the brand context file. this is the single document that makes every prompt in the session produce output calibrated to the specific customer, not generic AI copy.

if you don't have a brand context file yet, build one before you run this workflow. it contains five sections:

the customer avatar at the behavioral level (what they do at 2pm when they're struggling, what they've tried, what they've given up on)

the product mechanism (what it does and why, in one paragraph, no marketing language)

the visual identity (what the character looks like, what their environment looks like)

the tone reference (3 to 5 real testimonials or forum posts that sound the way you want the scripts to sound)

the performance log (which hooks worked and which got retired from previous batches, empty on day one)

the brief prompt:

read brand-context.md. set effort to maximum. we're building the weekly creative batch for [product]. reference the performance log for which angles to avoid and which to double down on.

that's it for setup. 90 seconds.

minute 5 to 10: generate 5 hook scripts (opus 4.8)

switch to opus 4.8 at maximum effort. this is the only step that needs the strongest model because the psychological depth of the hook determines everything.

write 5 UGC video scripts for [product] targeting [avatar]. each script: 12 to 15 seconds, first-person spoken dialogue.

rules:

hook (0-3s): a specific daily moment so real the viewer thinks "how did they know that about me." no product. use language from the customer avatar section verbatim.

confession (3-7s): what they've tried that didn't work. validates frustration. use the failed solutions from the research.

mechanism (7-11s): one sentence that teaches why the problem exists. earns credibility through education.

result (11-15s): specific and understated. "i noticed i was getting to 4pm without reaching for a third coffee."

five scripts in about two minutes of generation time. read each one aloud. if any line sounds like a brand wrote it instead of a real person recorded it on their phone, flag it for rewrite.

DM me "UGC" at @rubiinov if you want to see these scripts applied to your product.

minute 10 to 15: generate the character and product visuals (higgsfield MCP)

connect higgsfield MCP if you haven't already (settings, connectors, add custom connector, mcp.higgsfield.ai/mcp, authenticate, two minutes, one time only).

character generation (chatGPT images 2.0 model within higgsfield):

the critical rule: start with the outcome, not the product.

create a consistent character for our ad creative: [describe the avatar, not a model, a real person, what they look like on a regular morning, natural energy, not posed].

generate 4 reference images:

morning kitchen, natural light

post-workout, gym bag visible

home office, mid-afternoon

weekend morning, relaxed

save these four images. they're the consistency anchor for the entire batch.

product b-roll (same model):

generate 6 product scenes:

product on kitchen counter, slightly out of focus, hand reaching

product in gym bag, casual

product on desk, character in background

product held casually, not the focus

product with morning routine items

product mid-use

all: natural lighting, candid energy. product visible but never the hero.

total time for character + b-roll: about 3 to 4 minutes.

minute 15 to 22: animate into video (seedance 2.0)

this is where statics become footage. for each of the 5 scripts, animate the character reference image with full behavioral direction.

animate [character reference image] into a 12-second video based on script #1:

0-3s: character at kitchen counter, back slightly to camera, turns as if noticing the phone. natural. not performing.

3-7s: character facing camera, arms loosely crossed, leaning on counter. the body language of someone explaining to a friend.

7-11s: character picks up the product casually, looks at it briefly, sets it down. confidence of someone who already knows.

11-15s: small smile. not big. the kind you make when you remember something you forgot to say.

camera: slight handheld drift, not shaky, alive.

audio: ambient kitchen sounds. coffee maker. no music.

output: 9:16 and 4:5.

run this for all 5 scripts. each animation takes about a minute. with 3 environment variants per script (morning, afternoon, evening), you get 15 videos.

minute 22 to 25: extract statics and scale copy

static extraction: pull key frames from each animated video. the strongest single frame from each becomes a static ad for carousel and feed testing. 5 statics from 5 scripts.

copy variations (switch to sonnet 5):

take script #1's hook direction and generate 5 primary text variations, 5 headlines, and 5 descriptions across these formats: testimonial, problem-agitation, mechanism-education, result-forward, comparison.

run this for the top 3 winning hook directions. 15 primary texts, 15 headlines, 15 descriptions. paired with the 15 videos and 5 statics, you have a full creative matrix.

minute 25 to 27: organize and queue for upload

the output from the session:

5 hook scripts

15 animated videos (5 scripts x 3 environments)

5 statics from key frames

45 copy variations (15 primary text + 15 headlines + 15 descriptions)

9:16 and 4:5 aspect ratios for every video

organize by hook direction. each hook direction gets its own ad set in a broad-targeting CBO at $25/day. 72-hour testing window before touching anything.

total production time: 27 minutes.

total cost: claude subscription.

total assets: 30 (15 videos + 5 statics + 10 copy-paired variants).

the formats that work on meta cold traffic right now

not all AI ad formats perform equally. here's what's converting in the supplement category right now based on 600+ assets tested:

UGC testimonial (highest volume, most consistent): the hook/confession/mechanism/result arc described above. the workhorse. this format produces the most consistent 3-second view rates and the most predictable CTR. 60 to 70% of every batch should be this format.

problem-agitation static: a single enhanced image of the character in a relatable moment (staring at a laptop, rubbing their eyes, looking at a cluttered supplement shelf) with a one-line overlay that names the problem. no product visible. the static equivalent of the hook section. works for prospecting audiences who need the problem identified before they'll engage with a solution.

before/after transformation: two images, one showing the "before" state and one showing the "after" state, both featuring the same character. the before image uses cooler tones, lower energy, visible fatigue markers. the after image uses warmer tones, natural confidence, subtle environmental upgrade. works for audiences who are already problem-aware and need to see the outcome.

product-in-life candid: the product sitting in a real environment (kitchen counter, gym bag, desk) with no person and no text overlay. shot with natural lighting and slight depth-of-field blur. works as a retargeting creative for people who've already seen the UGC testimonial and need one more touchpoint before clicking.

mechanism educator: a 15-second video where the character explains one specific fact about why the problem exists. no product mention in the first 10 seconds. the product enters at the end as the logical conclusion. works for colder audiences who need education before they'll trust a product claim.

the creative matrix from a single 27-minute session covers all five formats. the algorithm tests all of them simultaneously and tells you which format your specific audience responds to.

the compounding effect: why week 4 looks nothing like week 1

the brand context file's performance log is what makes this system get better every session.

after batch one: you add which hook directions produced the highest 3-second view rates, which characters resonated, which result beats drove clicks.

after batch two: you add which formats outperformed (UGC testimonial vs. problem-agitation static vs. before/after), which environments produced the strongest completion rates, which copy formats earned the best CTR.

by batch four: opus 4.8 is writing scripts against a performance log that contains a month of market data specific to your customer. the scripts aren't guessing anymore. they're written against evidence.

the quality jump from batch one to batch four is not marginal. it's the difference between an operator running a creative process and an operator running a creative intelligence system.

what this replaces in your current workflow

if you're currently running a traditional UGC production workflow:

$4,200 to $7,500/month in creator fees replaced by a claude subscription

2 to 3 week turnaround replaced by 27 minutes

12 assets per month replaced by 120+

one round of revisions replaced by unlimited re-prompting

one creative direction per batch replaced by five simultaneous directions

guessing at angles replaced by testing at volume and letting the data decide

the operators who switched to this workflow in 2026 saw an average 2.2x improvement in ROAS within 60 days. the improvement comes from one thing: creative testing velocity producing enough data for the algorithm to optimize against.

the 27-minute session is the unlock. the compounding data asset is the moat.

if you want us to run this exact 27-minute production workflow on your brand and produce your first 30-asset batch today, DM me "UGC" and we'll make it happen

($20k+ days achieved through brands scaling from zero with this exact method, fully done-for-you)

rubinov

Links

• mcp.higgsfield.ai/mcp
• brand-context.md/
• mcp.higgsfield.ai/mcp