All tools
Open source Agent skills

higgsfield.ultimate-skills

Prompt-to-cinematic-video workflows.

  • Shell
  • Claude Code

01 · What it does

The problem it solves

AI video ads drift: the face, the garment and the colour grade change between shots, and picking the wrong model wastes credits. This is an 11-file reference library that pins down Seedance 2.0’s hard input and output limits, maps each Higgsfield image model to what it is good at, and supplies an Identity Lock block you paste into every video prompt so one locked hero still drives every shot. A worked 15-second sportswear example shows the method end to end.

Who it's for

Anyone building short-form video ads on Higgsfield whose characters change between shots, and Claude Code users who want the craft notes rather than working automation.

02 · Setup

Before you start

  • Node.js and Claude Code
  • A free Higgsfield account at higgsfield.ai
  • Playwright MCP, if you want browser automation

Clone

git clone https://github.com/ca-who-codes/higgsfield.ultimate-skills

Copy the reference library by hand

mkdir -p .claude/skills && cp -R reference .claude/skills/_reference

Add Playwright MCP

claude mcp add playwright npx '@playwright/mcp@latest'

03 · Step by step

Getting it running

  1. Clone the repo

    Pull down the reference library. The whole repo is about 4,500 words of markdown.

    git clone https://github.com/ca-who-codes/higgsfield.ultimate-skills
  2. Place the reference library yourself

    Do the one step the installer cannot: copy the reference directory to the shared knowledge folder the skills are meant to read. See the gotchas — install.sh currently fails.

    mkdir -p .claude/skills && cp -R reference .claude/skills/_reference
  3. Add Playwright MCP if you want browser control

    The docs assume Claude drives higgsfield.ai through a browser. Log in to the site once the Playwright window opens.

    claude mcp add playwright npx '@playwright/mcp@latest'
  4. Pick your image model

    The model guide maps each model to its URL path and strengths. The stated defaults are Soul 2.0 for people and UGC, Nano Banana Pro for product and 4K detail.

  5. Generate one hero still and lock it

    Iterate on a single still until it is right, then stop re-rolling. That one image becomes the reference for every shot in the spot.

  6. Paste the Identity Lock block into every video prompt

    The consistency file gives the block verbatim: preserve face, wardrobe and colours from the reference image, and name explicitly what is allowed to move. Describe identity in nouns, not vibes.

  7. Cut the spot from short shots in post

    Run three or four short generations from the same still and stitch them, rather than asking for one long take. Add logo, captions and voiceover in an editor, since Seedance does not render text reliably.

04 · What you get

Capabilities

Seedance 2.0 spec sheet

Hard limits in one file — up to 9 images, 3 videos, 3 audio clips, 12 files total, 4–15s of 720p output with sound.

Model routing guide

A table of image models with URL paths, what each is best at, and when to avoid it.

Character consistency playbook

The largest doc: a copy-paste Identity Lock block plus seven levers that keep faces and palette from drifting.

Camera, lighting and sound references

Three files of shot vocabulary, lighting setups and sound-design phrasing to drop into prompts.

2-second hook framework

A menu of opening hooks with the exact prompt phrasing for each.

Worked example

A full 15-second spot rebuilt using the consistency method.

05 · Before you rely on it

Limitations, honestly

Every tool has edges. These are the ones that will cost you time if nobody tells you first.

  • The published repo has no automation or skills directory, so none of the 19 slash commands the README lists exist in the clone.
  • install.sh exits 1 on the first copy and installs nothing, including the reference library — do that step by hand.
  • CLAUDE.md is a 0-byte empty file even though the README describes it as holding defaults.
  • Browser automation is described throughout, but no automation skill files are present to run it.
  • Model names, URL paths and Seedance limits are a point-in-time snapshot that goes stale whenever Higgsfield changes its lineup.
  • This is prose for a human or agent to read, not code — nothing to test and no CI. The README states MIT but there is no LICENSE file.

Read the source

Everything above came from the repository. Read it, fork it, or open an issue if something here is wrong.

github.com