how to turn claude + higgsfield MCP into a $10k+/day AI creative MACHINE

the creative production workflow i'm about to break down produces 40+ ad variants per week for my supplement brand.

near-zero per-asset cost. same-day production. no creators. no editors. no two-week turnarounds. and the output reads as real UGC because the production methodology is built around making the generated footage pass the same pattern-matching test that real phone footage passes.

this is not a "cool AI tool" walkthrough. this is the exact production system behind $10k+ days on meta, broken into the specific steps, the specific models, and the specific prompts that produce each component.

the production stack: what does what

sonnet 5 in cowork mode: competitive research and creative strategy

sonnet 5 is anthropic's most agentic sonnet. launched june 30, 2026. near-opus performance at a third of the cost. 1M context window. adaptive thinking on by default.

in cowork mode, sonnet 5 scrapes the meta ad library to analyze competitor creative. it's looking for: which hooks are running longest, which creative formats dominate, which angles are being tested across multiple brands, and which landing page types are being used.

sonnet 5 also scrapes your own meta account data or analyzes your current ads, landing pages, brand docs, customer reviews, and offer docs that you feed into the conversation.

chatGPT images 2.0 (via higgsfield MCP): product enhancement and static generation

the critical workflow insight that most operators miss: do not start with the product.

if you're selling peptides, don't start with a plain vial on a white background. nobody buys a peptide because the bottle looks nice. start with an image that shows the outcome the customer actually wants.

use a reference image of someone looking leaner. recovering after a workout. checking noticeably improved body composition in the mirror. then use the image generation model to enhance that image. make the lighting more cinematic. improve composition. add subtle environmental storytelling.

for a static ad, you're done at this point.

all of this runs through the higgsfield MCP inside claude.

seedance 2.0 (via higgsfield MCP): video animation

if you want a video ad, take the enhanced image into seedance 2.0 and animate it.

the behavioral direction is everything. the generation model doesn't imagine specificity. you have to tell it exactly what the character does.

for the peptide example: have the athlete finish a set. wipe sweat off their face. flex in the mirror. walk toward the camera. show subtle progress over time. keep the movement natural and believable.

the goal is not flashy AI effects. the goal is making the transformation feel real enough that someone scrolling actually stops and imagines themselves getting the same result.

the full production workflow: start to finish

step 1: competitive research (sonnet 5 in cowork, 45 minutes)

sonnet 5 scrapes the meta ad library for your brand and top 5 competitors. analyzes which hooks are winning, which formats dominate, which angles are saturated, which gaps exist. also scrapes reddit, amazon reviews, and youtube comments for fresh customer language. outputs a strategic creative brief with 8-10 prioritized angles.

step 2: hook script production (opus 4.8, 30 minutes)

opus 4.8 at maximum effort writes 5 UGC scripts using the hook/confession/mechanism/result arc. each script is 12-15 seconds of first-person spoken dialogue built on customer language from the research. not marketing copy. voice note energy.

step 3: outcome image selection and enhancement (chatGPT images 2.0 via higgsfield, 20 minutes)

select or generate the reference image. start with the outcome, not the product. enhance with cinematic lighting, environmental storytelling, and compositional refinement. generate 4 character reference images across different contexts for consistency.

step 4: product b-roll generation (chatGPT images 2.0 via higgsfield, 15 minutes)

generate 6 product scenes. natural lighting. candid energy. product visible but never the hero.

step 5: video animation (seedance 2.0 via higgsfield, 30 minutes)

animate each script-character combination with second-by-second behavioral direction, handheld camera drift, and ambient audio spec. output 9:16 and 4:5 for every asset.

step 6: static extraction (5 minutes)

pull key frames from the animated videos for carousel and static testing.

step 7: copy variation scaling (sonnet 5, 20 minutes)

take the winning hook directions and use sonnet 5 to generate 25-50 primary text, headline, and description variations.

total production time: approximately 2.5 to 3 hours for a full weekly batch. one operator. no creative team.

the claude skill that runs the entire workflow

skill name: weekly-creative-batch-production-v1

description: manages the full weekly creative production cycle for AI meta ads. orchestrates the research phase (sonnet 5 cowork), script phase (opus 4.8), visual generation phase (higgsfield MCP), animation phase (seedance 2.0), and copy variation phase (sonnet 5). maintains brand context file continuity across sessions and feeds performance data from previous batches into each new cycle.

trigger: when the user asks to produce a creative batch, run the weekly production cycle, generate new ads, or build creative for meta.

pre-production check:

• load the brand context file

• check for performance data from the previous batch

• if performance data exists, update the brand context file with: which angles won, which to retire, which directions to explore further

production sequence:

1. research phase: run the competitive research workflow using cowork mode

2. script phase: switch to opus 4.8 at maximum effort. generate 5 scripts

3. visual phase: generate character references (4 images) and product b-roll (6 scenes) through higgsfield MCP

4. animation phase: animate each script-character combination through seedance 2.0

5. static phase: identify key frames from each animation for static extraction

6. copy phase: switch to sonnet 5. generate 5 primary text, 5 headline, and 5 description variations per winning angle

post-production output:

• organized asset list: 15 videos, 5 statics, 25-50 copy variations

• campaign setup brief: broad CBO, $25/day per ad set, 72-hour test window

• measurement framework: the three metrics to watch and the decision rules for scaling vs retiring

the outcome-first principle: why it changes everything

the single insight that produces the biggest quality jump in AI-generated creative:

start with the transformation image, not the product image.

when you give a generation model a product shot and ask it to build an ad, the model centers the product. the output looks like product photography with a person nearby. that reads as an ad immediately.

when you give a generation model a transformation image and ask it to enhance and animate it, the model centers the person. the product enters the frame naturally as part of their routine. that reads as footage.

this distinction is the difference between AI creative that hits a 25% 3-second view rate and AI creative that hits a 51% 3-second view rate.

not a better prompt. not a better model. just the right starting image.

what $10k+/day actually requires in creative infrastructure

$10k/day on meta at a healthy ROAS requires approximately 20 to 30 active winning creatives at any given time, rotating on different fatigue cycles.

a brand producing 12 assets per month cannot sustain $10k/day because the creative dies faster than it's replaced.

a brand producing 40 to 60 assets per week has permanent creative depth. winners scale. fatigued assets rotate out. fresh creative enters before the old creative dies.

the gap between 12 assets per month and 60 assets per week is the gap between a $3k/day brand wondering why ROAS is flat and a $10k/day brand compounding every week.

if you want us to deploy this full claude + higgsfield MCP creative production system on your supplement brand and produce your first 40-asset batch in 30 days, 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