What the Higgsfield prompt skill actually encodes

A prompt library, not a model or an API client

The subject of this repository is a specific commercial product. Higgsfield AI is described as a cinematic video and image generation platform, and the project description positions the repository as a Claude AI skill for cinematic Higgsfield AI prompts, covering 32 sub-skills and 18 templates.

What the skill actually does is stated plainly: it transforms natural language requests into production-ready Higgsfield prompts. The input is what someone wants to see, the output is a prompt that names things the platform recognises. There is no inference here, no model weights, and no service call. The repository description makes the same point about the project itself, and the dependency file says it in the most concrete terms possible, since the skill content is Markdown and JSON and everything else in the repository tooling is pure standard library.

So who is it for? People who already use Higgsfield AI and want to write better prompts for it, and specifically people who are working on longer or more consistent productions where a single lucky prompt does not scale. The topics attached to the repository point the same way, with tags for ai-cinema, ai-filmmaking, ai-video, character-acting, cinema-studio, prompt-engineering, seedance, soul-id, motion-control and video-generation.

The size of the thing is the other fact worth stating up front. This is not a short prompt cheat sheet. It is a documentation tree with a spec snapshot dated 2026-09-26, a discipline framework, a failure catalogue, a production benchmarks file, an annotated template library and a set of scripts, all at version 3.40.0.

MCSLA is the spine the sub-skills hang from

The formula that gives the library its shape is MCSLA, expanded in the README as Model, Camera, Subject, Look, Action. Every prompt it teaches is assembled from those five slots, and the model slot is the one that changes most, because the sub-skill set covers model selection across Kling 3.0, Kling 3.0 Omni, Kling 3.0 Motion Control, Veo 3.1, Wan, Seedance 2.5, Seedance 2.0, FLUX 3 Video, Minimax Hailuo and Higgsfield DoP, among others.

The rest of the slot vocabulary is equally specific. Camera controls and motion presets are named the way the platform recognises them, rather than described in free text. The sub-skills for Seedance 2.0 list named techniques such as Intent over Precision, a Genre Router, an I2V Gate, Anti-Slop, Physics Language, SCELA audio, and prompt modes for reference-based work, continuation, expand shot, edit shot and transformation, along with a Continuation Prompt Formula and an Iteration Rule.

There is a second structural idea that is not part of MCSLA but governs how the sub-skills cooperate. DISCIPLINE.md is described as a cross-cutting framework of 9 named discipline patterns arranged in 3-3-3 tier symmetry, covering prompt construction, model selection and iteration discipline, and it is applied across the sub-skills rather than inside one of them.

For someone evaluating this, the practical question is whether that layering transfers to a different platform. It does not, and the repository does not pretend otherwise. The vocabulary is platform vocabulary: named presets, Cinema Studio panels, Soul ID, MCSLA slots. The structure is reusable, the words are not.

The four Seedance 2.5 modes and the reference roles

The most detailed part of the library is the Seedance 2.5 omni-reference director, and it is worth reading closely because it shows what careful prompt engineering for a video model looks like in practice.

There are four generation modes: t2v, omni_reference, video_edit and video_extension. The reference system is explicit about roles, using @Image, @Video and @Audio markers and stating what each one may and may not carry. On top of that sits a five-step multi-reference workflow, 30-second staging with declared end states, timestamp pacing, and bracket syntax for music, sound effects, dialogue and subtitles.

Frame control is handled two ways, either through the platform's own start_image and end_image in omni_reference mode or declared inside the prompt, plus multi-keyframe control, storyboard grids, and a distinction between coarse and fine blockout rendering. There is also a routing table for choosing between 2.0 and 2.5.

The same four-mode surface appears again under the Cinema Studio 4.0 sub-skill, which is described as a CLI workflow, and that entry carries a caveat that matters for anyone planning to rely on it: it is schema-documented, not yet field-rated. In other words, the repository records the shape of the 4.0 surface without claiming to have tested it against real generations.

The Seedance 2.0 sub-skill adds a second layer of guidance, with a template library, a reference-based routing table and an Anti-Slop section, and the AI-VFX pipeline sub-skill goes further into production mechanics, covering model routing per asset class for faces, creatures, clothing and locations, a plain-grey sheet law, a face-lock crop, a size-ref frame for scale between two subjects, location batching selected on light, and a four-batch stop rule with an i2v fallback.

For teams already working in Seedance 2.0 or 2.5, this is where most of the practical value sits, because it encodes decisions about which lane to use and when to stop, rather than just how to phrase a shot.

Four versions of Cinema Studio, four different surfaces

A recurring source of friction in this platform, judging from the library, is that the tool interface changes. The sub-skills cover Cinema Studio 2.5, 3.0, 3.5 and 4.0, and they describe genuinely different things.

Cinema Studio 2.5 covers Soul Cast AI actors, built-in color grading, 3D Mode using Gaussian Splatting, Grid Generation, resolution settings, a frame extraction loop, object and person insertion, per-character emotions, clustering, a five-view location reference sheet, reference sheet types for motion, outfit and palette, and an elements system with a library surface of five source tabs across six element categories.

Cinema Studio 3.0 is narrower in reach and wider in capability. It is described as a Business and Team plan feature, and it brings native dual-channel stereo audio, smart shot control, a 15 second maximum duration, 7 genres, @ reference patterns and Soul Cast 3.0.

Cinema Studio 3.5 is a different interface again, described as a three-pill main UI of Genre, Style and Camera, with a style settings panel of 8 colour palettes, 6 lighting options and 9 camera moveset styles plus a manual style mode, and a camera settings four-axis panel of 3 camera bodies, 5 lenses, 5 focal lengths including a new 75mm, and 3 aperture options. It also has an image mode with a four-model picker.

Keeping four interface descriptions in one repository is the price of writing about a product in motion. It also means a reader has to know which version their plan exposes, and the 3.0 guidance applies to Business and Team plans only, so a reader on a different plan can follow the prompt craft and ignore the surface details.

Consistency machinery: Soul ID, acting profiles and failure catalogues

Character consistency gets the most attention in the library, and it is treated as a discipline problem rather than a prompting trick. There is Soul ID character consistency guidance with character sheet creation, an identity versus motion separation rule described as hard, for keeping a character consistent across shots, and the character sheet work inside the Hell Grind feature-film pipeline, which the README attributes to Higgsfield's open-sourced 95-minute feature system.

That pipeline section is the most demanding part of the repository. It covers headless character sheets, mask-composited point edits, location sheets with anchors and one light logic, a per-scene GEO SPATIAL LAYOUT block, a position-fixing first second, dialogue construction, a ban dictionary, a 10 to 15 iteration rule, and crowd, giant and threshold solutions. There is also a separate production-benchmarks file describing the Hell Grind 90-minute Cannes feature reference, per-character iteration anchors and acceptance-rate calibration.

The acting sub-skill is a performance system in its own right, built around objective, obstacle, tactics, beats and subtext, plus listening markers, body, status and proxemics, mandatory eye life, a 150 to 220 word acting master profile with a per-scene rewrite, a locked voice prompt, a 15-symptom atlas of bad acting and a 0 to 5 performance scale.

Failure diagnosis is treated the same way. FAILURE-MODES.md documents 15 named render failures for Seedance, each with symptom, mechanism and counter, and the library also carries a shared negative constraints reference with categorised artifacts and prevention phrases, a Kling 3.0 Motion Control failure diagnostic, and a physics rendering resolution decision matrix that routes 480p, 720p and 1080p across Seedance 2.0 and Cinema Studio 3.x.

That diagnostic framing is the most transferable idea in the repository. Whether or not you use Higgsfield, symptom-mechanism-counter is a better structure for prompt iteration than guess and retry.

The repository as a working tree with its own tooling

The top level listing shows a project with infrastructure, not just prose. Alongside README.md, SKILL.md, CHANGELOG.md, CONTRIBUTING.md, DISCIPLINE.md, INDEX.md and LICENSE there are skills/, specs/, templates/, docs/, evals/, tests/, scripts/, db/, workspace/ and assets/ directories, plus several standalone reference files such as model-guide.md, image-models.md, photodump-presets.md, prompt-examples.md, production-benchmarks.md and vocab.md.

The dependency file is where the practical detail sits, and it is short:

# Python dependencies for repo tooling (not needed to use the skill itself).
# The skill content is Markdown/JSON; only the PDF user-guide generator needs a
# third-party package. Everything else (scripts/validate.py, scripts/higgsfield_memory.py,
# scripts/seedance_lint.py) is pure stdlib.

Three things follow from that. There is no installation step for using the skill, because the skill is text. The tooling named there, a validator, a memory script and a seedance linter, runs on the standard library alone, which also means the repository holds no client for the Higgsfield platform and no dependency on a vendor SDK. And a test suite exists, with pytest listed as a development and CI requirement that is not needed to use or validate the skill.

The .claude/ directory and the platform badge suggest the skill is packaged for Claude Cowork and Claude Code, and the .github/ directory and .markdownlint.json point to a repository with formatting and review conventions. For a first real use, the honest workflow is: read INDEX.md and SKILL.md, install the skill in your Claude environment, run the validator over your own prompts, and use the failure catalogues to name what went wrong before rewriting the prompt.

Licence, cost and the parts you cannot verify

The project is MIT licensed, with the licence file at the repository root, and the version badge in the README reads 3.40.0 against a spec snapshot dated 2026-09-26. The last push was on 2026-09-27, and the three most recent releases, v3.38.0, v3.39.0 and v3.40.0, all landed on that same day. A repository can be current and still churny, and three releases inside one day is churn. The CHANGELOG.md file in the tree is the place to read what those releases actually changed.

The MIT licence covers the repository text, which is the useful part, but it does not change the commercial terms of the platform the prompts target. Two of those appear in the library itself. The repurposing sub-skill, covering Shorts Studio, Clipify, Virality Predictor and Video Analysis, is described with an explicit paid and free split, and the Canvas sub-skill is described with a build-free, generate-paid cost model. Cinema Studio 3.0 is a Business or Team plan feature, so the 3.0 guidance applies to a paid tier.

That is where the honest limitation sits. The most elaborate part of the repository, the parts covering the 4.0 surface, the AI-VFX pipeline and the feature-film pipeline, is written against a platform that changes quickly and, in the case of 4.0, is explicitly not yet field-rated. There is also no field-verification harness in the repository, despite the evals/ directory, that the README describes as evidence of production quality.

So the appropriate use is as a vocabulary and a checklist rather than as truth. The named presets, the failure modes and the decision rules will save you a lot of trial and error. The specific claims about what a given platform version will do should be re-checked against the live product before you build a production workflow on top of them.

Editorial conclusion

Adopt this skill if you generate video on Higgsfield AI and have found that your prompts drift, that you cannot keep a character consistent across shots, or that you cannot tell a failed render from a bad prompt. Do not adopt it as a pipeline integration, because the repository tooling is standard library only and does not talk to the platform. Verify first which Cinema Studio version your plan exposes, since the 3.0 surface is described as Business and Team plan only and the 4.0 surface is explicitly schema-documented rather than field-rated, then check the CHANGELOG.md in the repository for how far the recent release churn has moved the guidance.

Frequently asked questions

Official sources

1. License: MIT
2. OSideMedia/higgsfield-ai-prompt-skill on GitHub
3. Project website
4. README
5. Releases