Turn Prompts Into Content With Claude Code + Higgsfield

Go to higgsfield.ai/cli. Higgsfield gives you three setup commands.

Run them in a fresh project folder:

npm install -g @higgsfield/cli
higgsfield auth login
npx skills add higgsfield-ai/skills

The second command opens a browser so you can sign in with your Higgsfield account. A free account works for testing. Paid plans and extra credits are useful once you start generating more assets.

When you run the skills command, choose Claude Code when prompted. You can install the skill for the current project or globally. For a first test, project-level install is cleaner because it keeps the experiment contained.

Pro tip: if setup gets confusing, ask Claude Code to help you install it. It can check whether Node and npm are installed, run higgsfield --help, and explain any terminal errors.

Once the CLI is installed, open Claude Code in the same folder.

Ask it to inspect the Higgsfield CLI page and the local command:

Look at https://higgsfield.ai/cli and inspect the local Higgsfield CLI.
Confirm the install worked, show me the command pattern, and list the available image models.
Do not generate anything yet.

Claude Code should check the CLI before spending credits.

This is where the workflow starts to click. Claude Code can see the Higgsfield command, understand the available model names, and decide which model IDs to use in later commands.

If Claude Code lists video models too, tighten the instruction and tell it to use image models only for this test.

Now give Claude Code one image prompt and ask it to run a real comparison.

Use this prompt:

Use the Higgsfield CLI to generate this prompt with six different image models:

"A retro 1970s travel photo of Mount Fuji at sunrise, soft film grain, warm colors, vintage postcard composition."

Create a `higgsfield-model-test` folder.
Save each output with the model name.
Create `comparison.md` with the model, command, output file, and quick notes.
Use image models only.
Download the final image files locally when each generation finishes.

That last line matters. In testing, the first generation returned a hosted image link. If that happens, just tell Claude Code to download the image when it is done.

The useful part is not just that Higgsfield can generate an image. The useful part is that Claude Code can run the commands, wait for results, save files, and keep the comparison organized.

Use the saved outputs to compare framing, texture, color, and realism across models before choosing a direction.

After the six images are saved, ask Claude Code to review the folder and summarize the differences.

Use a prompt like this:

Review the images in `higgsfield-model-test`.
Update `comparison.md` with:
- the strongest image overall
- the best image for social
- the best image for a newsletter hero
- what each model did well
- what prompt change we should try next

Then run one more focused pass. For example, if one model nailed the color but another handled composition better, ask Claude Code to combine those notes into a tighter prompt and test two or three finalists.

You can also use this as a cost-control workflow. Test rough concepts on cheaper models first, then send the strongest prompt to a higher-quality model once the direction is clear.