Higgsfield AI Prompt Skill

Higgsfield AI Prompt Skill

Language rule: Reply in whatever language the user writes in.

HARD RULES — pre-delivery checklist

These rules apply to every Higgsfield response. They are written as a pre-delivery checklist the agent runs before sending the response, not as prohibitions stated and then forgotten. The failure mode they prevent is plausibility-over-verification — producing a response that looks correct because the agent's training data knows the rough shape of Higgsfield work, rather than because the agent actually read the skill files and verified the platform's ground truth.

Before delivering any Higgsfield response, confirm in this order:

1. Routing line present. First line of response names which sub-skills you routed to (e.g. "Routing to higgsfield-prompt + higgsfield-camera for an Atmosphere push-in"). One line, then the work. Missing routing line = response is incomplete; add it.

2. Routed sub-skills opened and read in this conversation. Match the user's ask to the routing table below, open the matching sub-skill files with the read tool, and READ them. Root SKILL.md and skills/higgsfield-prompt/SKILL.md are mandatory at minimum on any prompt request. Grepped snippets do not satisfy this rule. Full reads do. If your only access to root SKILL.md or skills/higgsfield-prompt/SKILL.md in this conversation came from grep results, you have not satisfied this rule — open the file. Platform vocabulary, preset names, and model parameters must come from the files because this platform's lineup changes between releases.

3. Named vocabulary verified, not invented. Camera preset names, motion preset names, model names, CLI flag forms, and MCP tool parameter names all come from the skill files or from verification. For model parameters, enums, and durations, verify against specs/model-specs.yaml first — it is generated from a dated models_explore snapshot (see snapshot_date inside the file); if the snapshot is stale (>30 days), verify live instead (higgsfield model get <model> for CLI param schemas; models_explore for MCP). If you found yourself thinking "this flag probably looks like X" or "this preset is probably called Y" — stop. Read the file or run the verification command. Plausibility is not validity. Do not substitute generic video-prompt vocabulary for named Higgsfield presets; do not invent model versions, camera presets, or motion-preset names. If the user names one you don't see in the skill files, say so and ask for clarification.

4. MCSLA structure intact on video prompts. Model · Camera · Subject · Look · Action. Five layers, every video prompt, unless the user explicitly opted out.

5. Shared negative constraints appended. Pull positive-phrasing prevention phrases from skills/shared/negative-constraints.md. Do not paraphrase from training; use the exact phrasing from the file. (Kling 3.0 prefers positive phrasing over negations; using negation-form constraints when the file says positive is a fidelity miss.)

6. Preflight surfaced when applicable. If execution intent is signaled (CLI / MCP / bundled-skills mentioned) AND a video-class or high-cost model is named OR a budget concern is named, surface the two-step preflight (model get / models_explore for schema, then cost estimate). See skills/higgsfield-stack/SKILL.md § Preflight discipline.

7. Aspect ratio is an enum, not a free-form value. Check the model's allowed ratios against specs/model-specs.yaml before writing them into the header; if the snapshot is stale (>30 days), verify live via schema (models_explore / model get). Example of why this matters: Seedance 2.0 supports native 21:9, Kling 3.0 does not. Anamorphic / 2.35:1 / 2.39:1 are style register vocabulary for the Look line, not output ratios. See vocab.md § Aspect Ratio: output spec vs. style register. In a multi-shot Seedance sequence whose location plates carry a baked lens, whether the optics words still belong in the video prompt (Style Prefix included) is OPEN — one studio drops them, Hell Grind keeps the look in both (skills/higgsfield-seedance/SKILL.md § Bake it into the asset; skills/shared/house-rulings.md P2-6).

8. Prompt under 200 words — short-form regime only. Soft cap from MCSLA section. Going over is a signal you're padding rather than locking — tighten. Regime exception: block-scaffold production prompts (skills/higgsfield-seedance/SKILL.md § Official Prompt Architecture) replace the word cap with structural lint — harvested production Seedance briefs run 218–2,059-word medians depending on register [FIELD — 13-project community harvest, 2026-07-18]. The cap governs single-shot MCSLA prompts; a block-scaffold prompt over 200 words is not a rule-8 violation.

If any of items 1–8 are missing or unverified, the response is incomplete. Complete them before sending, not after.

What Is Higgsfield?

Higgsfield is a cinematic AI video and image generation platform built for filmmakers and creators. Unlike single-model tools, Higgsfield hosts multiple generation engines on one platform — Kling 3.0/3.0 Omni/3.0 Motion Control, Google Veo 3.1/3.1 Lite, Wan 2.7/2.6/2.5, Seedance 2.5/2.0/Pro, FLUX 3 Video, Minimax Hailuo 2.3/02, Higgsfield DoP (Lite/Standard/Turbo) for video; Soul 2.0, Soul Cinema Preview, Soul Cast, Nano Banana Pro/2, Kling Image 3.0/Omni, Seedream 5.0 Pro/Lite/Flash + 4.5, GPT Image 2.0 / 2.5, Flux 2/Kontext for images — plus a library of 100+ named Motion Presets, a Soul ID