Claude Opus 5.5 + Higgsfield: I Built an AI UGC Reel Engine for $7 a Reel

Kristoffer · October 1, 2026 · 19 min read

The reel at the top of this video is a UGC ad. The guy in it doesn't exist. The product he's holding doesn't exist either. And the whole thing was written, filmed, edited and posted to Instagram without me ever touching a camera.

Here's why that matters. Brands pay real creators around $200 for a single video like this, and one AI UGC tool, Arcads, is reportedly making around $15 million a year selling this kind of content to brands.

So I had Claude Opus 5.5 build my own version: a marketing engine that creates an AI avatar, invents a product, films a UGC reel of that avatar promoting it, then plans a whole week of reels and posts them automatically. It costs about $7 per reel instead of $200.

This post walks through the whole build: the architecture, the API keys, the build, the first avatar, product and reel, the redesign with Mobbin, and a full week of reels going live on Instagram.

Why UGC, and why you need so much of it

Building the product isn't the hard part anymore, because AI can do that for you. The hard part is getting people to actually see it.

The content that sells on TikTok and Instagram right now isn't polished ads. It's UGC, short for user-generated content: a normal-looking person holding the product and talking to their phone camera like they're telling a friend about it.

The problem is volume. One video won't do anything on its own. You need new videos every day, because you never know which one will take off. At $200 a video, posting once a day costs you about $6,000 a month.

How the marketing engine works

Think of the engine as a tiny marketing team with three members:

Role Tool Job
The brain Claude Opus 5.5 Comes up with the product, angles, scripts, captions and posting plan
The camera crew Higgsfield Generates the avatars, product photos and talking videos
The social media manager Blotato Posts the finished reels to Instagram or TikTok

How they're connected is the important part:

Step by step, the engine:

  1. Creates an AI avatar, or picks one from a library of avatars you've already made.
  2. Creates a physical product with AI, like a skincare product, with real-looking product photos.
  3. Writes a UGC script with three beats: a hook, the product demo, and the reaction.
  4. Films each beat as a talking video of the avatar using the product and speaking the lines.
  5. Stitches the clips together, adds captions, and outputs a finished vertical reel.
  6. Posts it to Instagram or TikTok, or plans out a whole week at once.

Why this is worth building

There are three criteria almost every viral app follows, including tools like Arcads:

Arcads went from its first customers to millions a year with a team of fewer than 10 people. You only need to build a slightly better version for yourself.

Step 1: Get your Higgsfield and Blotato API keys

Before Opus writes a single line, you need two keys. This takes about two minutes.

Higgsfield

  1. Go to open.higgsfield.ai, Higgsfield's developer platform, and open the dashboard.
  2. Top up your API balance. This is the part that confuses people: the Higgsfield API has its own balance. Paying for a Higgsfield plan on the website doesn't give you API credits. You top it up separately and pay only for what you generate. I had about $55 on mine, which was plenty for this whole build.
  3. Click Create API key, name it (I used "build with ai"), and leave expiration off.
  4. Copy the key straight away. It's only shown once.

The dashboard also lists the models the API gives you, like Seedance, Kling and Marketing Studio. We use a few of them.

Blotato

Blotato is a social media scheduler. I picked it because its API lets the app post to TikTok, Instagram and a bunch of other platforms, all with one key.

  1. Go to Settings, then the API tab.
  2. Create a key and click Copy API key.
  3. Connect your Instagram (or TikTok) account inside Blotato. You only do this once.

API access is only on Blotato's paid plans. I only connected Instagram for this build, but the app handles TikTok the exact same way.

Step 2: Set up Claude Code with Opus 5.5

  1. Create a completely empty folder called marketing-engine.
  2. Open a terminal inside it and start Claude Code.
  3. Check the model. Mine was already on Opus 5.5 with the 1 million token context window, so no switching was needed. I left effort on medium, the default.

Then type /mcp. Higgsfield's MCP was sitting right there, but switched off on purpose. An MCP only works while you're sitting there chatting with Claude, and I wanted an app that runs on its own from a button. That only works through the API.

Step 3: The build prompt

The biggest mistake beginners make is opening an AI tool and typing "build me a marketing app." The most important thing before generating a single screen is a proper spec: a clear description of every feature, every screen, and what "finished" actually means.

For this app, the hard part isn't the screens, it's the pipeline:

So the prompt tells Opus exactly:

Skip that structuring detail and your app will break. Here's the exact prompt. Paste it into Claude Code:

Build a complete local web app called "Marketing Engine," a ready-to-use template that automates short-form marketing with AI. It creates AI avatars, creates AI-generated physical products (or uses a real product the user uploads), generates 20-30 second vertical UGC reels of an avatar promoting the product, and schedules those reels to TikTok and Instagram. Use React, TypeScript, Vite, and Tailwind CSS for the frontend, and a local Node.js + TypeScript server for everything else. The app must run locally with npm install and npm run dev. Include an .env.example. Never expose any key to the browser.

The product must feel like one coherent app, not a collection of unrelated demo pages.

Architecture - three services, and these rules are strict:

A. The brain is Claude Opus 5.5 running through the locally installed Claude Code CLI in headless mode. The server calls "claude -p" as a child process with model claude-opus-5-5 and JSON output, using my existing Claude Code login. Do NOT use the Anthropic API, the Anthropic SDK, or an ANTHROPIC_API_KEY. Keep the model name in one config value. Opus 5.5 writes all text: product concepts, avatar descriptions, image and video prompts, UGC scripts, captions, hashtags, and the weekly content plan. Every response must be strict JSON, validated against a schema before use, with retries on malformed output, a timeout, and a clear error if Claude Code isn't installed or logged in.

B. All images and videos come from the Higgsfield REST API only (base URL https://api.higgsfield.ai, header "Authorization: Key KEY_ID:KEY_SECRET", env vars HIGGSFIELD_API_KEY_ID and HIGGSFIELD_API_KEY_SECRET). Do NOT use the Higgsfield MCP server, the Higgsfield CLI, or the higgsfield.ai website. Read the current docs at docs.higgsfield.ai and the model pages at open.higgsfield.ai before writing the client. Every generation is async: submit, get a request_id, poll GET /requests/{request_id}/status with backoff (2s rising to 10s, with jitter) until completed, failed, nsfw, or canceled. Respect the concurrency limit (queue jobs, at most 4 at a time). Download every result immediately into a local media library, because Higgsfield only keeps outputs for about 7 days. Upload local files through POST /files/generate-upload-url. Use these models: avatars - higgsfield-ai/soul/v2/standard (aspect_ratio 9:16, batch_size 4, 1080p); products - workflows/product-shots; scene keyframes of the avatar holding or using the product - marketing-studio/image with the avatar and product images in image_urls, aspect_ratio 9:16; talking video clips - bytedance/seedance-2.0/reference-to-video with the keyframe, avatar and product in image_urls, aspect_ratio 9:16, generate_audio true, 720p, duration 4-15 seconds, and the exact spoken line written in the prompt. Keep every model path and default setting in one config file.

C. All posting goes through the Blotato REST API only (base URL https://backend.blotato.com/v2, header "blotato-api-key", env var BLOTATO_API_KEY). Do NOT use Blotato's MCP server or its built-in AI video tools. Read the current docs at help.blotato.com (start with help.blotato.com/llms.txt) before writing the client. Load connected accounts with GET /users/me/accounts. Upload finished videos using Blotato's documented media upload for local files, then create posts with POST /posts, and poll GET /posts/{postSubmissionId} no faster than every 10 seconds until scheduled, published, or failed. Never resend a create request blindly, since Blotato has no idempotency key; store every postSubmissionId.

Required product areas:

1. Setup - A first-run setup screen that checks, with a clear green or red status for each: Claude Code installed and logged in, Higgsfield keys valid, Blotato key valid with at least one TikTok or Instagram account connected, and ffmpeg installed. Explain how to fix each red item in one sentence. If keys are missing, the app still runs in a clearly labeled demo mode with sample media, and never pretends a demo result is real.

2. Avatar Library - A grid of saved avatars with name, look, and vibe. "Create avatar": the user describes who they want (age range, gender, style, setting, vibe, or just "surprise me"); Opus 5.5 turns that into a realistic, phone-camera, UGC-style Soul 2 prompt; Higgsfield returns 4 options; the user picks one, names it, and saves it. Allow regenerating, deleting, and marking favorites. Saved avatars are reused as reference images so the same person appears in every clip.

3. Products - Two ways to add a product. "Generate with AI": the user types a rough idea or a category, Opus 5.5 creates a full product concept (name, what it is, who it's for, 3 key benefits, packaging design, price point, and the tone of voice for ads), then Higgsfield creates product shots; the user picks the best and saves it. "Use my product": upload real photos and a description. Every product has a detail page with its images and concept, all editable.

4. Reel Studio - Pick a product, an avatar, a platform, and an angle (problem/solution, "TikTok made me buy it", unboxing, honest review, before/after, or let Opus choose). Opus 5.5 writes a 3-beat UGC script: a hook (about 5 seconds), a product demo (about 8-12 seconds), and a reaction and call to action (about 5 seconds). Each beat has the exact spoken line, the scene and camera direction, and on-screen caption text. Spoken lines must sound like a real person talking to their phone, not like an ad, and must fit the beat's duration. The user can edit the script before generating. Before anything is generated, show the estimated Higgsfield cost and ask for confirmation. Then generate a keyframe and a talking clip for each beat, showing live progress per beat. Stitch the clips with ffmpeg into one 1080x1920 MP4, burn in bold, readable, TikTok-style captions timed to each beat, and normalize the audio. Let the user play the reel, regenerate a single beat without redoing the others, and download it.

5. Caption and post copy - For every reel, Opus 5.5 writes a separate TikTok caption and Instagram caption with hashtags, editable before posting. The captions must never claim that an AI-generated product can be purchased unless the user has marked the product as real and added a link.

6. Scheduler - Choose which connected TikTok and Instagram accounts to post to, then post now, schedule for an exact date and time in the user's time zone, or use Blotato's next free slot. TikTok posts always set isAiGenerated to true, with privacy level, comments, duet, and stitch settings exposed. Instagram posts go out as Reels with share-to-feed on. Show a queue and a week calendar of everything scheduled, with live status from Blotato (scheduled, published, failed), the public post link once published, and clear error messages with a retry option.

7. Content Week autopilot - "Plan my week": pick a product, one or more avatars, how many reels (1-14), and posting times. Opus 5.5 plans a different angle and hook for each reel so they don't feel repetitive, shows the plan and the total estimated cost for approval, then generates every reel in a background queue and schedules each one through Blotato. The queue survives a server restart and can be paused or cancelled.

8. Spend and history - Track every Higgsfield request with its model, cost estimate, status, and time. Show spend today, this week, and per reel, plus a user-set daily budget that blocks new generations once it's reached. Keep a history of every reel and post.

9. Settings - Light, dark, and system theme. Time zone. Default platforms, captions style, reel length, and video resolution. Reset demo data. Export and import all data as JSON.

Design requirements: Premium, modern creative-studio design, not a generic admin dashboard. It should feel like a tool a founder wants to open every morning. Reels and avatars are always shown in 9:16 frames. Strong hierarchy, excellent spacing, thoughtful empty states, subtle motion, clear progress for long-running jobs, and consistent components. Accessible contrast and keyboard-friendly controls. Avoid excessive gradients, random glassmorphism, or unnecessary visual clutter. Create a distinctive visual identity for Marketing Engine.

Engineering requirements: Store data in a local SQLite database and media files in a local media folder. Use reusable components and clean project organization, with separate, typed clients for Claude Code, Higgsfield, and Blotato. Never spend Higgsfield credits or post to social media during automated tests: mock all three services in tests. Handle failed, nsfw, and timed-out generations with clear messages and a retry. No broken buttons or placeholder interactions for required features. Avoid console errors. Include a README with setup instructions, where to get each key, an architecture summary, estimated costs, and a checklist of implemented features.

Work autonomously. First inspect the empty project and read the Higgsfield and Blotato docs, create a concise implementation plan, then build the full app. Test the app yourself in demo mode before declaring completion. Do not spend real credits or publish real posts while testing. Do not stop after creating a mockup or partial prototype. Continue until the required functionality is implemented and the project runs successfully.

The parts that do the heavy lifting:

I ran the session with bypass permissions on, so Opus could work as long as it needed without stopping for approval. It started by inspecting the project and my computer, read the Higgsfield and Blotato docs, and checked what was installed: Node, npm, ffmpeg, Claude Code itself and git. That's exactly what you want, because the app depends on all of them.

Step 4: What Opus built in 1 hour and 6 minutes

The whole build took 1 hour, 6 minutes and 37 seconds. The summary is worth reading:

First run

I pasted my keys into the .env file and ran npm run dev. The server confirmed the brain is Opus 5.5 via Claude Code, and both the Higgsfield and Blotato keys are set.

The app opens on a setup screen with four checks, all green:

Then the studio itself:

I switched to light mode and unticked TikTok as a default platform, since I'd only connected Instagram. One bug: on the Avatars page in light mode, the "Let Opus cast someone" text was almost invisible. Noted for later.

Step 5: Cast the AI avatar

Click Create avatar. Instead of one big text box, you get a little casting form: age range, gender, style, setting and vibe. I cast him like a real creator: a man, 25-34, professional, wearing a suit in a skyscraper office overlooking Miami, with a calm, authoritative vibe.

Here's what happens behind the scenes:

  1. Click Write the prompt and Opus 5.5 turns the description into a detailed image prompt in about 6 seconds. It starts with "Candid iPhone front camera selfie," so the result looks like a real phone photo, not a glossy AI portrait. You can still edit it.
  2. The prompt goes to Higgsfield's Soul 2 model, which is built for realistic people, asking for four options at once in vertical 1080p.
  3. Before spending anything, the app shows the cost: 2 cents for all four.

About 30 seconds later the four options were back, and they honestly look like real phone selfies: same suit, same office, same Miami view. Option three looked like a noticeably younger guy, so I went with option two. Opus suggested a name, look and vibe for him, so he became Marcus, saved in the library. Every reel from now on can use the exact same person, which is what makes a brand feel consistent. Don't like any of the options? Click Regenerate options for four new ones.

Step 6: Invent the product

In real life you'd upload your own product photos here. To show what's possible, I let AI invent the entire product. I clicked Generate with AI and typed:

a skincare product for men who hate complicated routines

A few seconds later Opus 5.5 came back with a full product concept:

That's a product I'd believe exists, and "Wash. Swipe. Go." is a genuinely good tagline.

For the product shots you can pick up to three Higgsfield preset styles. I chose Balloon Headline Trio, Accept or Decline, and Frosted Toggles. Estimated cost: 70 cents. The app renders a clean packshot first, then builds the three styled shots from it.

The label actually comes out readable - "Dayshift", "One-Step Face Stick" and "Wash. Swipe. Go." - which AI usually gets wrong. One miss: I asked for a single stick and the main packshot showed two (front and back). I kept the clean packshot as the main shot and saved the product.

Notice the "Real product people can buy" switch is off. That means captions will never claim you can buy it, because it doesn't exist. The real cost for all four images was about 35 cents.

Step 7: Film the first UGC reel

In Reel Studio, Dayshift and Marcus were already picked since they're the only ones we have. Platform: Instagram. Length: 24 seconds. Angle: let Opus choose. Click Write script.

Opus picked the problem-to-solution angle and titled it "Marcus - desert or oil slick":

Beat Length Setting Line
Hook 5s At his desk in a high-rise Miami office, tilting his face to the window light to show a shiny T-zone "Suit's fine. My face is either a desert or an oil slick."
Demo 12s Office washroom "I'm not doing five steps. Wash. Twist this up. Swipe. Rub it in. 20 seconds, dries matte, so no shine in meetings, and it lives in my gym bag."
Reaction 7s Back at his desk "That's it. That's the whole routine. Would you actually use this? Tell me below."

Every beat comes with its own camera direction and on-screen caption. It sounds like a real guy talking to his phone, not an ad, so I didn't change a word.

Before spending anything, the app shows the estimate: $8.04. An app that can spend money on its own needs to ask first. Then, for each beat:

  1. Higgsfield Marketing Studio creates a keyframe: a still of Marcus with the product in the scene Opus described.
  2. That image goes into Seedance 2.0, which turns it into a video where he moves and speaks the line in his own voice.
  3. ffmpeg stitches the three clips together, burns in the captions, and levels the audio.

Progress for every beat shows live. The reel was ready in about 8 minutes.

Honest review of reel #1

What worked: the voice sounds natural, it's the same Marcus in all three clips, and the Dayshift label is readable when he holds it up.

What didn't:

So I'd still watch every reel before it goes out. Real cost for this reel: $7.50, compared to about $200 from a human creator.

Step 8: Redesign the app with the Mobbin MCP

The engine works, but the app looked generic and still had that invisible-text bug. That matters because this is a tool you'll open every single day, and if it looks cheap, you start trusting its output less.

So instead of telling Opus to vaguely "make it look better", I gave it real references from Mobbin, a massive curated library of interfaces from real, successful apps. Mobbin has an MCP, a plug that lets Claude search it directly. Using an MCP is fine here, because it only helps Opus design the app while I'm working with it; it's not part of the engine that runs on its own.

To add it, run this in your terminal (or type it inside Claude Code with a ! in front):

claude mcp add mobbin --scope user --transport http https://api.mobbin.com/mcp

Type /mcp and you should see Mobbin connected with three tools. Then send the redesign prompt:

Use the Mobbin MCP to research strong, modern product interfaces relevant to Marketing Engine: creative studio and video editing tools, AI generation interfaces with galleries and option pickers, avatar and character libraries, social media schedulers and content calendars, job queues and progress states, usage and billing dashboards, and onboarding and account connection flows.

Do not copy one product screen-for-screen. Extract the strongest recurring patterns and use them to improve Marketing Engine as an original product.

First, create design-research.md with: the products and flows you examined, the patterns you kept, the patterns you rejected and why, and a complete Marketing Engine design system (visual direction, colors, typography, spacing, radii, shadows, icons, motion, how 9:16 media is framed, progress and job states, status colors for scheduled, published, and failed, empty states, mobile behavior, and accessibility rules).

Then implement the redesign across the entire app. Preserve and test all existing functionality. Do not reduce features. Make light and dark mode equally polished. Do not spend Higgsfield credits or post to social media while testing; use demo mode. Do not declare completion until the redesign is consistent across every screen and the project runs without console errors.

The key line is "Do not copy one product screen-for-screen." You want the strongest recurring patterns across many apps, not a clone of one.

It finished in 44 minutes and 55 seconds:

Switching to dark mode, the difference is night and day. It's not perfect - there's an empty gray box in the appearance settings, and in dark mode the slider on Content Week has no visible handle - but it finally feels like a real product.

Step 9: Plan and post a full week of reels

This is the part the whole engine was built for. In Content Week I picked Dayshift, Marcus, and seven reels starting tomorrow. It suggested two posting times, 9:00 AM and 6:00 PM; I removed 9:00 AM, so it's one reel a day at 6:00 PM for seven days to my Instagram.

Opus 5.5 planned the whole week, and they're actually different videos, not the same one seven times:

Every reel has a different setting too: an office bathroom, a kitchen counter, a balcony, a car, and a hotel room.

The estimated total was $56.26, with a warning that it's more than what's left of today's budget, so any reels that don't fit today wait in the queue and continue after midnight. I approved it, and the engine ran on its own, working on three reels at a time: script, keyframes, then the talking clips.

The budget guard in action

We went over the $25 daily budget and the app blocked new generations until midnight. I wanted to finish the week the same day, so I raised the budget to $50 and the reels kept coming. We hit the cap again at $50, and I raised it once more to $70.

Day 3, the gym bag before-and-after, came out fully done:

"Open my gym bag. Everything's coated in lotion again. Cool. Love that. So I switched to this. Twist, swipe, rub it in. Done. 20 seconds, tops. And look, no shine. Same face at 4 as at 9. Bag's dry. Face isn't shiny. Honestly, that's all I wanted. Would you try this or stick with soap?"

Live on Instagram

The week view showed reels ready, one scheduled for Monday, and one already published. On Instagram, there it was: a real reel, live, starring a person who doesn't exist, selling a product that doesn't exist, and I didn't film, edit or post a thing.

The caption Opus wrote talks about the routine and asks a question, but never says you can buy it, exactly as we asked. In Blotato, the next reel was already scheduled for the same 6:00 PM slot on Monday (Blotato shows it in your local time).

Average cost across the week: about $7.20 per reel.

Recap

Stage Tool Result
API keys Higgsfield + Blotato Separate Higgsfield API balance, Blotato paid plan, Instagram connected
Build Claude Code + Opus 5.5 Full app in 1h 6m, 41 passing tests, crash-safe generations, budget guard
Avatar Opus 5.5 + Higgsfield Soul 2 Marcus, four options for 2 cents
Product Opus 5.5 + Higgsfield Dayshift Face Stick with readable labels, about 35 cents
First reel Marketing Studio + Seedance 2.0 + ffmpeg 24-second, three-beat reel in 8 minutes for $7.50
Redesign Mobbin MCP 70 screens studied, cool neutral palette with one orange accent
Content week Opus 5.5 + Blotato Seven different reels planned, generated, scheduled and posted at about $7.20 each

We started with an empty folder and ended with a marketing engine that has an avatar library, AI products, UGC reels with real voices, a spend tracker with a budget guard, and a full week of reels posting to Instagram on their own - without writing a single line of code.

AI can still get a hand or a label wrong, like we saw in the first reel, so review every reel before it goes out.

Frequently asked questions

How much does an AI UGC reel cost to make with Higgsfield?

In this build, a 24-second, three-beat reel cost $7.50, and the average across a full week of seven reels was about $7.20. Avatars cost 2 cents for four options, and four product images cost about 35 cents. A human UGC creator typically charges around $200 per video.

Does a Higgsfield subscription include API credits?

No. The Higgsfield API has its own balance on open.higgsfield.ai, separate from any plan on the main website. You top it up and pay only for what you generate.

Why use the Higgsfield API instead of the Higgsfield MCP?

An MCP only works while you're chatting with Claude. The API lets the app call Higgsfield on its own from a button or a schedule, which is what lets the engine generate and post a whole week of reels without you.

Do I need an Anthropic API key?

Not for this setup. The app uses Claude Code on your own computer as its brain, which runs on your normal Claude subscription.

Is it okay to post ads for a product that doesn't exist?

This build marks the product as not purchasable, so the captions never claim anyone can buy it. For real campaigns, upload your own product photos and use the same engine to promote something people can actually buy.

If you want this exact marketing engine, every prompt and the finished project, it's all inside BuildWithAI, along with the source code for every other app from my videos and the private founders community.