A Claude Image Generation Skill vs Higgsfield: What the Demo Shows
By Matt Cullerton, founder of Mavric and Apso
A TikTok from Nate Herk (@nateherkai) promotes a Claude skill called "Generate" that, in the Whisper transcript, "turns Claude into its own Higgs field." He most likely means Higgsfield, the multi-model image and video generation platform, though the audio alone cannot confirm he said the company name. I found no public listing, repository or documentation for a skill named "Generate," so its existence rests on the video. Higgsfield already connects to Claude through its own CLI and MCP server. The skill's claimed difference is cost control, a spending cap and cheapest-provider routing, though the video shows these only as mockups and the model list on screen is dated.
What is Higgsfield?
Higgsfield AI is a platform for generating images and video across many third-party models. Its CLI page says the tool lets you "Generate images and videos across 30+ models without leaving your terminal," with "30+ models including Soul, Cinema Studio, Flux, Seedream, Kling, Minimax Hailuo, Veo, and more."
The video's transcript renders the name as "Higgs field," the physics term. Nothing in the video touches particle physics, and every capability described matches the platform, so Higgsfield is the likely reading. The only evidence is the Whisper output plus context.
Does Higgsfield already work with Claude?
Yes. The CLI page says you can "Run Higgsfield from the command line in Claude Code, Cursor, or any MCP-compatible CLI agent," and lists compatibility with "Claude (web, Cowork, and Claude Code), OpenClaw, Hermes Agent, and NemoClaw." Setup is:
npm i -g @higgsfield/clihiggsfield auth loginnpx skills add higgsfield-ai/skills
Higgsfield's Claude integration page says "Claude connects with Higgsfield MCP and orchestrates the leading video models (Soul, Veo, Kling, Seedance, and more)."
What does the Claude skill claim to add?
Spending control. The demo prompt on screen reads: "Create me 10 different ads for this product using three different image models, and don't spend more than $3." The narration says Claude "will even find the cheapest provider for each one." Other frames show a toast reading "New spending limits set to $5" and the caption "GETS SAVED locally."
Higgsfield's Claude page uses a different rule: "Claude picks the best model for the shot automatically, or you can name one." On pricing it says "Paid usage runs on a credit system, priced by model, duration, and resolution." It does not mention a dollar cap, cheapest-provider routing or local saving.
The video shows those features only as motion graphics. No frame shows the skill running, and it is distributed by DMing "GENERATE" on Instagram.
Which image models are in the picker, and are they current?
The model picker shows Nano Banana, Nano Banana 2, GPT Image 1.5, Wan 2.6, Wan 2.7, Wan 2.7 Pro and Seedream v4.5. Two entries are out of date.
GPT Image 1.5 is being retired. OpenAI's deprecations page lists gpt-image-1.5 as deprecated on June 2, 2026, with shutdown on December 1, 2026 and gpt-image-2 as the replacement. The video posted September 23, 2026.
Wan 2.6 carries a "NEW" badge while Wan 2.7 and Wan 2.7 Pro sit beneath it. Wan 2.6 is Alibaba's December 2025 release, and Wan 2.7 now supersedes it.
The frame captioned "THE CHEAPEST" shows a different picker filtered to "Top Search Models": claude-opus-4-6-search, grok-4.20-beta1, gpt-5.2-search, gemini-3-flash-grounding, gemini-3-pro-grounding, gpt-5.1-search and claude-sonnet-4-6-search. These are language models with web search. None generates images, so that footage does not illustrate cheap image routing.
Can you customize a Claude skill?
Yes. The video says "because it's a Claude skill, you can customize the entire thing around how you work. You can add new models, you can set different spending limits, or you can create your own saved visual styles."
Anthropic's Agent Skills announcement, from October 16, 2025, describes skills as "folders that include instructions, scripts, and resources that Claude can load when needed." They work in the Claude apps for Pro, Max, Team and Enterprise users, in Claude Code, and through the API. A folder of instructions and scripts can be edited, so the customization claim holds.
The only frame showing a real agent at work is Claude Code adding a three-line theme toggle to web/src/pages/ProjectPage.tsx, which is unrelated to image generation.
What are the products in the "generated" ads?
They are real brands: Ormonde Jayne's Sakura perfume, ARKIVE by Adam Reed hair products including The Future Youth Treatment Mask (180 mL), and an UPTIME drink. The video does not say whether these were generated by the skill.
When was this recorded?
Shortly before Anthropic merged Cowork into Claude. The composer in the video shows a "Chat | Cowork" toggle. VentureBeat reported that on September 16, 2026 Anthropic announced the merge, and that "The new design removes that explicit Chat-versus-Cowork choice."
The book on the shelf behind Herk is his own, Becoming AI Native.
Bottom line
You can already run Higgsfield's 30-plus models from Claude through its official CLI and MCP server. The "Generate" skill, which has no public listing I could find, pitches budget caps and cheapest-provider routing, which Higgsfield's Claude page does not offer, but the video shows those features only as mockups. If you use any picker like this one, swap GPT Image 1.5 for gpt-image-2 before December 1, 2026.
Sources
- Higgsfield CLI - model catalog, install steps and Claude support
- Higgsfield Claude AI Video Generator - MCP integration and pricing model
- OpenAI API deprecations - GPT Image 1.5 shutdown date
- Anthropic: Introducing Agent Skills - what skills are and where they run
- VentureBeat on the Cowork merge - Chat and Cowork merge
- Ormonde Jayne Sakura - the perfume in the video
- Becoming AI Native on Amazon - Nate Herk's book
- ARKIVE The Future Youth Treatment Mask at Boots - the hair product in the video
Published September 23, 2026. Based on a TikTok by @nateherkai. Full video notes (markdown)