17:33
Higgsfield
The video positions Higgsfield Supercomputer and the Higgsfield CLI (run inside Claude Code) as opposite ends of a technical-difficulty spectrum for building AI product commercials, and argues that choosing between them comes down to cost (shared Higgsfield credits vs. paid Claude + Higgsfield subscriptions) and capabilities (an easy all-in-one browser chat vs. a more technical, local-file-based setup that can be remote-controlled from a phone).
Three-tier spectrum of technical difficulty: Higgsfield Supercomputer (easiest, browser-based) → Higgsfield MCP (middle, Claude desktop app) → Higgsfield CLI (most technical, Claude Code)
The two biggest differences between the setups are framed as cost and capabilities
Higgsfield Supercomputer merges multiple LLMs (Gemini, Claude, ChatGPT) with Higgsfield's creative tools, skills, connectors, and a persistent memory in one interface
Demo workflow: upload a product reference image, a Pinterest style screenshot, and a brand identity guide, then ask Supercomputer to extract visual identity and use the GPT Image 2 Director skill to write prompts and generate visuals
Recommends enabling 'ask before generation' so credits aren't spent automatically without review
More context about the product/brand yields more accurate results; a sparse made-up reference set caused inconsistencies like product-to-hand scale issues
Credit usage splits into image credits (generation) and text credits (usage of Gemini Orchestrator, Claude, or ChatGPT in the background)
Supercomputer has a Community tab for one-click skill installs plus the ability to upload custom skills
Thomas shares two of his own free custom skills — the C-Dance 2 multi-shot prompter and the Image Prompt Helper — usable inside Supercomputer or the CLI
The C-Dance 2 multi-shot prompter turns a set of still images into a commercial video by asking structured questions about mood, shot count, and shot flow
A rushed, uncarefully-read prompt produced a flawed C-Dance video, illustrated live, and the mistake was fixed by sending the video back to Supercomputer with written critique since it has Gemini's video understanding built in
Connectors tab includes integrations such as a Telegram bot for chatting with Supercomputer remotely
Setting up Higgsfield CLI requires installing Claude Code, then running install commands from Higgsfield's CLI/MCP tab, and requires paid Claude and Higgsfield plans
The CLI workflow operates on local files organized into per-product folders, with Claude Code executing multi-step generation instructions via the Higgsfield CLI
Thomas built and shares a free 'Higgsfield download wrapper' that auto-uploads/downloads files between local folders and the CLI session
Claude Code's remote control feature lets you text instructions to an active session from a smartphone (via the Claude app, same login) while the desktop keeps running, as long as it stays on and connected to the internet
Disclaimers given: back up files/computer if new to Claude Code, and disconnect remote sessions when finished for safety
Higgsfield Supercomputer — An all-in-one browser interface that combines multiple LLMs (Gemini, Claude, ChatGPT) with Higgsfield's creative tools, skills, connectors, and a persistent memory. Apply: Open the Higgsfield homepage, click Supercomputer, and chat with it (uploading a product reference image, brand guide, and style references) to extract visual identity and generate product visuals.
Higgsfield MCP — The middle-tier setup that connects Higgsfield directly inside the Claude desktop app, covered in Thomas's earlier video. Apply: Use it inside the Claude desktop app to run a product-commercial pipeline starting from a single reference image.
Higgsfield CLI — The most technical setup, connecting Higgsfield directly to Claude Code via the command line for local-file-based generation. Apply: Install Claude Code, install the Higgsfield CLI using the commands shown on Higgsfield's CLI/MCP tab, then instruct Claude Code to generate and download images/videos using files in a local project folder.
GPT Image 2 Director — A built-in Supercomputer skill that extracts visual identity from uploaded reference material (Pinterest screenshots, brand guides) and writes generation prompts. Apply: Ask Supercomputer to use GPT Image 2 Director to turn extracted visual identity into prompts before generating product images.
Ask-before-generation setting — A Supercomputer option that pauses before spending credits on an image generation instead of auto-firing. Apply: Enable it so you can review and check a box to approve generation before credits are spent.
Image credits / text credits — Higgsfield's two-part credit accounting, where image credits track generation cost and text credits track use of the underlying LLMs (Gemini Orchestrator, Claude, ChatGPT) in the background. Apply: Watch both credit meters in Supercomputer's usage panel to track spend across chat reasoning versus actual image generation.
Community skills tab — A library inside Supercomputer of user-made skills installable with one click, plus support for uploading custom skills. Apply: Browse the Community tab to install skills such as Thomas's C-Dance multi-shot prompter and Image Prompt Helper for free, or upload your own.
C-Dance 2 Multi-shot Prompter — Thomas's custom skill that turns a set of generated still images into a multi-shot commercial video by asking structured questions about mood, shot count, and shot flow. Apply: Add generated images to the task, invoke the skill in chat, and answer its questions (energy/mood, number of shots, shot sequence like dropper, splash, portrait, hero) to produce the commercial video.
Image Prompt Helper — Another of Thomas's custom skills for writing detailed image-generation prompts. Apply: Install it from the description and invoke it inside Supercomputer or the CLI to write prompts from extracted visual identity for use with a product reference image.
Iterative video critique loop — A correction technique enabled by Supercomputer's built-in Gemini video understanding, where a flawed generated video is sent back with written critique to produce an improved regeneration. Apply: After a bad C-Dance generation, resend the video along with specific notes (shot order, tempo, sound effects, hero-shot animation) and ask Supercomputer to retry the multi-shot prompt.
Connectors tab / Telegram bot — A Supercomputer feature offering integrations, including a Telegram bot for chatting with Supercomputer remotely. Apply: Set up the Telegram connector from the Connectors tab to discuss and generate with Supercomputer over Telegram.
Higgsfield download wrapper — A tool built and shared by Thomas that automatically uploads and downloads files between a local folder and a Higgsfield CLI session. Apply: Install it (linked in the description) so generated images and videos are automatically saved into your local project folder without manual upload/download steps.
Remote control feature (Claude Code) — A Claude Code feature that lets you control an active session — including its Higgsfield CLI generation tasks — from a smartphone via the Claude mobile app using the same login. Apply: Start a session on your desktop, invoke the remote control command in Claude Code, then text further instructions from your phone while the computer stays on and connected, disconnecting the session when done.
The real dividing line between the two setups isn't raw capability but which credit/subscription structure you're willing to commit to — a single Higgsfield credit pool for Supercomputer versus stacked paid Claude Code plus Higgsfield subscriptions for the CLI
Because a complex skill like the C-Dance multi-shot prompter takes a detailed structured prompt, rushing through it doesn't just produce a worse result — it burns real credits on a wrong generation, as the video itself demonstrates
Supercomputer's built-in Gemini video understanding turns a failed generation into a conversational fix: you can resend the bad video with specific critique (shot order, tempo, sound effects, animation) instead of manually re-editing
The CLI setup's core practical edge over Supercomputer is that it works on local files, which combined with the download wrapper means finished assets land automatically in a project folder rather than staying inside the Higgsfield platform
Remote-controlling a local Claude Code + Higgsfield CLI session from a phone effectively turns a left-on home computer into an unattended generation-and-delivery pipeline ('just like an employee') for as long as it stays connected
«Now, during the past couple of weeks, I've been going deeper and deeper into the rabbit hole of Higgsfield in combination with Claude.»
— 00:22
«And then, on the most technical and scary side, we have the Higgsfield CLI that you can use directly inside of Claude Code.»
— 01:37
«To make it really simple for you, the biggest two differences between these setups is cost and capabilities.»
— 01:44
«the result we got back is actually a perfect example of why you should always always read through your prompts carefully because now I rushed a little bit in this section and didn't read through the prompt carefully.»
— 08:39
«here I want to give a disclaimer that if you're completely new to Cloud Code, watch some videos on YouTube of how to set it up in a safe way and always have backups of your files and your computer.»
— 11:58
«And then you can walk away from your computer and it will continue working for you just like an employee.»
— 15:46
Reception
Audience is enthusiastically engaged and appreciative, finding the content educational and inspiring, with constructive questions about tool choices and costs rather than criticism.
A transparent, demo-driven comparison that clearly maps the ease/cost/capability tradeoffs between Supercomputer and the CLI, but it deliberately avoids declaring a winner; the on-camera C-Dance misfire doubles as an unplanned cautionary example of how costly careless prompting can be with these credit-based tools.
17:33