How to Set Up Higgsfield with Claude Code to Produce Great Videos

In this week’s letter I told the story of a garden that arrived as a picture first. This is the workbench that made the picture.

Three tools, one pipeline. Claude Code is the cockpit — I talk to it in plain English and it runs the whole thing. Nano Banana (Google's image model) is the art department — it generates and edits stills while keeping the same face, place, or object consistent frame to frame. Higgsfield is the motion studio — it turns those stills into video, adds camera moves, and assembles the final cut. The magic isn't any one tool; it's that they hand off to each other through a single conversation, so I can go from idea to finished piece without leaving the keyboard — and stop to approve or redirect at every step.

What each tool actually does

Claude Code the cockpit

This is the orchestration layer. Instead of clicking around three different web apps, I describe what I want in natural language and Claude Code drives the other tools for me. It also carries the context that matters — the look I'm after, the project's rules, what's been tried already — so what comes out is pointed at the idea instead of generic.

Why it's the hub: it connects to both Nano Banana and Higgsfield through "connectors," so one conversation can generate an image, critique it, edit it, and animate it — with me steering the whole way.

Nano Banana the art department

Nano Banana is Google's image generation and editing model. Its superpower is consistency: it can keep the same character's face, the same backyard, or the same object intact while you change the scene, the lighting, or the composition around it. That's the thing most image tools get wrong, and the thing that matters most when you're building a world instead of a one-off.

What I use it for: concepting stills, art direction, and iterating. "Same yard, golden hour, wider shot." Each edit is a sentence, not a re-shoot.

Higgsfield the motion studio

Higgsfield takes still images and brings them to life. It does image-to-video (a still becomes a moving shot), camera moves and motion control (push-ins, pans, the kind of movement that reads as "produced"), character sheets for consistency across shots, upscaling to 2K/4K, and reframing so one piece ships in every aspect ratio a platform needs. It can also assemble multiple shots into a finished sequence.

What I use it for: turning locked stills into story videos and social cuts — the production step that used to require a crew.

How they hand off — the pipeline

1. Brief in Claude Code. I describe the piece I want. Because Claude carries the project's context, it starts on-target.
2. Generate stills with Nano Banana. Claude calls the image model to produce key frames. We iterate in plain English until the look is right — consistent across every frame.
3. Lock the keyframes. I pick the shots that work. This is a human decision — the taste gate.
4. Animate with Higgsfield. Each locked still becomes a motion clip. Camera moves and pacing get applied here.
5. Assemble & finish. Clips get stitched, on-screen type added, upscaled, and reframed for each platform (vertical for social, wide for everything else).
6. Deliver. Export, review, ship.

The whole loop can run in one sitting — and because exploring a direction is cheap, I can try five concepts before committing to one, instead of betting everything on a single idea.

Here's what that actually looks like on my screen — one conversation, three products handing off:

The session, verbatim shape — brief in plain English, tools called by name, a human approval between every step.

Setting it up (one time)

1. Install Claude Code — the command-line app. This is the driver's seat.
2. Connect the creative tools as MCP connectors: the Higgsfield connector and an image-generation connector (Nano Banana / Google's image model). Once connected, Claude Code can call them directly from the conversation.
3. Point Claude at your project's context — the look, the rules, the reference images — so everything it generates starts from your world, not a blank one.

Making a piece

1. Describe the deliverable in one message. E.g. "A 15-second story: a backyard garden growing out of a printed picture — warm morning light, four cedar beds under oaks." Claude turns that into a shot list.
2. Generate the stills. Ask for the key frames. Review them. Refine with plain-language edits — "the second one, but pull back and add the bench." Consistency holds across the edits.
3. Approve the frames. Nothing animates until you say a frame is right. This is where quality lives — kill the weak frames early.
4. Animate. Hand the approved stills to Higgsfield. Specify the motion — a slow push-in, a pan, a subtle drift.
5. Assemble. Stitch the clips, add any on-screen copy, set the pacing, upscale.
6. Reframe and export. Generate the platform variants (9:16, 1:1, 16:9) from the same master, then share.

Things that make it work

Why bother

• Speed. Idea to finished piece in a sitting, not a production cycle.
• Consistency. The same faces, places, and light across a whole body of work — the thing that usually breaks with AI tools.
• Cheap exploration. Try several directions before committing to one.
• The human stays in charge. Every step has an approval gate. The tools do the labor; the judgment stays with you.

What came off the line — the garden, imagined before it exists

That's the whole workbench. The picture of the garden — the one my wife is holding while the real beds wait for September — came off this exact line. The tools are new. The gift is old.

— Joe