how to cook VIRAL AI videos with sonnet 5 + higgsfield MCP

you type a brief into Claude.

90 seconds later, a finished 9:16 vertical clip lands in your chat. 4K resolution. platform-native pacing. AI-generated audio synced to the visual. ready to post on TikTok, Reels, or Shorts without opening a single editing tool.

that's not a pitch. that's what i'm running right now, and i'm going to hand you the exact setup, the exact prompts, and the Claude skills that make it work at a level where the output doesn't read as AI slop but as content that competes with what creators are manually producing.

the stack: Claude Sonnet 5 as the brain. Higgsfield MCP as the rendering engine. and a set of production prompts and skills that turn the raw capability into repeatable, quality-controlled output that you can run at volume without babysitting.

the setup (under 3 minutes, no code, no API keys)

step 1: connect Higgsfield MCP to Claude

open Claude on web, desktop, or mobile. go to Settings > Connectors. click "Add custom connector." name it "Higgsfield." paste the server URL: mcp.higgsfield.ai/mcp. save it, click Connect, sign into your Higgsfield account. done.

step 2: understand what's connected

once Higgsfield MCP is live, Claude has access to 30+ video and image generation models through five tools: generate_image, generate_video, create_character (Soul ID for identity consistency), get_generation_status, list_characters.

step 3: select Sonnet 5 as your model

Sonnet 5 launched June 30, 2026. it's now the default on Free and Pro plans. why Sonnet 5 specifically: it's the most agentic Sonnet Anthropic has shipped. it plans across steps, uses tools (including Higgsfield MCP) autonomously, and sustains focus on complex multi-step production tasks without stalling or losing context.

the production prompts (copy, paste, modify, run)

prompt 1: the UGC product ad (DTC brands, app founders)

you are a creative director producing a UGC-style product video for TikTok.

product: [paste your product URL or describe the product in 2-3 sentences]

target audience: [describe your buyer: age, platform behavior, what they care about]

the video must:

open with a hook that stops the scroll in the first 1.5 seconds (use a pattern interrupt, not a product shot)

feel like a real creator made it, not a brand (use the UGC preset in Higgsfield)

run 8-12 seconds total

be 9:16 vertical, 1080p minimum

include generated audio that matches the energy of the visual

hook approach: [choose one: reaction to discovering the product / POV: you just found this / the thing nobody told you about [product category] / i tested this for 30 days]

generate 3 variations using different models (Soul for the hero version, Kling 3.0 for the social cut, Seedance 2.0 for the most realistic option).

prompt 2: the surreal physics scroll-stopper (supplements, wellness, beauty)

generate a 5-second 9:16 vertical video using the surreal physics visual style.

the visual: [describe the impossible physical moment, e.g., "a glass jar of honey slowly inverting and the honey flowing upward against gravity, catching light, impossibly smooth and viscous"]

requirements:

the physics violation must be subtle enough to create ambiguity (the viewer should need 2-3 seconds to realize something is wrong)

lighting should be warm, natural, close-up

no text overlays

use Kling 3.0 or Cinema Studio

4K resolution, no watermark

generate 2 variations with slightly different physics violations (one gravity-based, one texture-based).

prompt 3: the freeze-to-motion viral format

i'm going to upload a product photo. generate a freeze-to-motion video where the still image comes alive over 5 seconds.

format: 9:16 vertical, 1080p, generated audio that matches the visual energy

model: use Nano Banana Pro for the initial still quality assessment, then Seedance 2.0 for the animation

the motion should feel organic and surprising, not like a simple parallax zoom.

the Claude skill that makes this repeatable at volume

a single prompt produces a single clip. to produce at volume, you need a Claude Skill: a reusable instruction set that Claude loads before every generation.

here's the Skill:

you are a video production director operating the Higgsfield MCP pipeline for a creator brand / DTC brand / AI app. your job is to produce platform-native short-form video content at volume while maintaining quality standards that compete with manually produced creator content.

production standards every video must pass before delivery:

HOOK TEST: the first 1.5 seconds must contain a visual that stops the scroll. if the opening does not pass this test, regenerate with a different hook approach before delivering.

PLATFORM NATIVE TEST: the video must feel like it belongs on the target platform's feed. this means 9:16 vertical, pacing that matches the platform's dominant content rhythm, and audio that sounds native rather than stock.

ANTI-SLOP TEST: if the output looks obviously AI-generated (uncanny valley faces, physics artifacts, text rendering errors, temporal inconsistency), flag it and regenerate using a different model before delivering. never deliver content that reads as AI slop.

MODEL SELECTION LOGIC:

photorealistic product shots and UGC: Soul or Seedance 2.0

surreal physics and impossible visuals: Kling 3.0 or Cinema Studio

freeze-to-motion and image animation: Seedance 2.0

fast social cuts and trend formats: Kling 3.0

cinematic and high-production: Cinema Studio or Veo 3.1

character consistency across clips: Soul with trained character ID

BATCH PRODUCTION: when generating a batch (5+ clips), vary the model, hook type, and visual approach across the batch. never produce 5 clips that look like they came from the same template.

VIRAL SCORING: after generating each clip, use Higgsfield's viral scoring tool to evaluate hook strength, retention risk, and viral potential. include the score in the delivery. flag any clip scoring below 70% for replacement.

delivery format: for each clip, include the clip, model used, viral score, recommended platform, and one-line hook copy suggestion for the caption.

the cost math

a single production session producing 10 clips:

Sonnet 5 tokens: approximately $0.50-$1.00

Higgsfield credits: approximately 10-30 credits

total cost per clip: under $5 including everything.

compare to: a human editor producing the same 10 clips from scratch would charge $500-$2,000.

why sonnet 5 specifically (not opus, not fable, not 4.6)

sonnet 5 over opus 4.8: Sonnet 5 at $2/M input tokens produces equivalent results for tool-coordinated production workflows at 60% lower cost.

sonnet 5 over fable 5: Fable 5 at $10/M is 5x Sonnet 5's introductory rate. use Fable 5 for strategic analysis. use Sonnet 5 for production execution.

sonnet 5 over sonnet 4.6: Sonnet 5 sustains multi-step agentic tasks measurably longer than 4.6. in a production session with 10+ Higgsfield renders, 4.6 would lose coherence by clip 7. Sonnet 5 maintains them through clip 15 without degradation.

if you want us to audit your brand and show you how we'd implement this exact Sonnet 5 + Higgsfield MCP production system to generate viral AI video content at scale over the next 90 days (or we keep running it until the view counts are there), DM me "COOK" and we'll make it happen

(200,000+ creator network running AI UGC campaigns across every major platform, fully done-for-you)

john

Links

• mcp.higgsfield.ai/mcp
• mcp.higgsfield.ai/mcp