Higgsfield Video Explainer

Higgsfield Video Explainer

Run the MCP video-explainer workflow through Higgsfield CLI. Lock one visual style key, write one narration line and one matching visual prompt per 10-second block, generate every voice take first, generate every clip second, then immediately assemble the ordered pairs with explainer_video.

Never use the monolithic video_explainer job in this skill.

MCP-to-CLI mapping

nano_banana_2 is the public CLI id for the Nano Banana Pro style-key model used by the MCP workflow.

Bootstrap

1. If higgsfield is unavailable, install it:

   bashCopy code
   curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh
   

2. If higgsfield account status fails, ask the user to run higgsfield auth login, then wait.

3. Inspect the live contracts before the first submission:

   bashCopy code
   higgsfield model get nano_banana_2
   higgsfield model get seed_audio
   higgsfield model get gemini_omni
   higgsfield model get explainer_video
   

Phase 0 — ask first

Collect choices in two separate turns, in this order. Never merge them.

Turn 1 — style only

Always load the live CMS catalog:

bashCopy code
higgsfield preset list video-explainer --json

Show the preset names with their thumbnail/video preview URLs. Say one short line asking the user to pick a preset, describe a custom style, or attach style-reference images, then end the turn. Do not ask production questions in the same turn. Choosing a style is mandatory; never choose silently unless the user explicitly says “you choose.”

Skip this turn only when the request already contains explainer preset id: <uuid>. Confirm that UUID exists in the live catalog and keep it for Phase 1.

Turn 2 — production settings

Only after style selection, collect every unresolved setting:

• Duration: one to ten whole minutes. N = duration_minutes × 6 fixed 10-second blocks.
• Narration language: English by default, but still offer the choice.
• Character: recurring mascot or faceless stylistic scenes. Always ask.
• Aspect: 16:9 by default or 9:16 vertical.
• Subtitles: off by default. Explain that subtitles cost 0.05 credit per voiced block. If enabled, make the user choose patrick, caveat, marker, or anton; never choose silently.

Every choice belongs to the user unless they explicitly delegate it.

Inputs

• Topic or personal/philosophical story.
• Optional local source documents; read/extract them before scripting. They are factual input, not generation media.
• Optional preset UUID, mutually exclusive with custom style-reference images.
• Optional style-reference images. Use only their rendering style and color grading; never copy their people, text, logos, or objects unless requested.
• Duration, language, character mode, aspect, and subtitle choice from Phase 0.

For local style donors, pass each path with a repeated --image. For a web image, download it locally first or use an existing uploaded media ID.

Hard rules

• Keep every visual strictly non-photorealistic. Repeat the same STYLE descriptor and non-realism negatives in every clip prompt.
• Keep all spoken words out of video generation. Clip audio is ambient sound or music only; no dialogue, lip-sync, or baked narration.
• Use exactly one narration take and one clip per labeled block. Block N audio always maps to Block N video.
• Attach the same style-key image to every clip.
• Write all image/video prompts in English. Only narration uses the selected language.
• Research real topics before scripting. Do not invent quotes, dates, numbers, or events.
• Assemble automatically in the same run. Returning loose clips is a failure.

Pipeline

Read references/prompts.md before Phases 1–3.

Phase R — research

For a real topic, use available web research tools and authoritative sources to verify enough facts for every block. Keep a short Sources list. Never script a factual explainer from memory alone.

For a personal story, skip web research and use only details supplied by the user. Invent nothing factual.

Phase 1 — create or resolve the style key

Write one reusable STYLE descriptor: medium, palette, line/fill behavior, texture/finish, then non-photorealistic, illustrated, not a photo, no live-action, no realism.

Selected CMS preset

Resolve the hidden style image into the active workspace:

bashCopy code
higgsfield preset resolve video-explainer "<preset UUID>" --json

Keep the returned media_id as STYLE_KEY_ID. Skip image generation: this imported media is the style key. Build the STYLE descriptor from the returned preset name plus the mandatory non-photoreal rules. Do not recreate a preset from its name.

The preset reference controls framing. If it conflicts with the aspect requested in Phase 0, stop and let the user choose rather than silently fighting the reference.

Custom style or reference images

Generate exactly one key image. Use the abstract swatch template from references/prompts.md, or its mascot variant when character mode is enabled. Repeat --image for every style donor:

bashCopy code
higgsfield generate create nano_banana_2 \
  --prompt "<style-key prompt>" \
  --aspect_ratio 16:9 \
  --resolution 2k \
  --wait \
  --json

Use 9:16 for vertical. Keep the completed image job UUID as STYLE_KEY_ID; later CLI generations can reuse a completed job UUID as an image reference.

Phase 2 — write narration

Write exactly N labeled narration blocks in the selected language:

textCopy code
Block 1
<line spoken over clip 1>
Block 2
<line spoken over clip 2>

• One line per block, usually 20–24 words and about 8–9 seconds.
• Keep every take under roughly 9.5 seconds.
• Use plain spoken text only: no timecodes, emotion cues, parentheticals, or stage directions.
• Spell numbers out.
• Use a concrete tone and never say “in this video.”
• For a topic, build from hook through payoff. For a personal story, preserve the user's details and protagonist.

Phase 3 — write matching video prompts

Write exactly N labeled English prompts using the template in references/prompts.md:

textCopy code
Block N
STYLE REFERENCE: Match the attached reference image EXACTLY. <same STYLE descriptor>
SCENE: <one scene and action matching Block N narration>
MOTION: <camera move and animation behavior>
AUDIO: <ambient SFX or music only; no voice, dialogue, or narration>
NEGATIVE: <style drift and realism bans; no lip-sync, captions, text, logos, or watermark>

For mascot mode, Block 1 greets by gesture with mouth closed, the final block waves a sign-off, and middle blocks use consistent cameos only when useful. For faceless mode, use stylistic scenes only. Keep one clear action per block.

Phase 4 — generate every voice take first

List the live voices, present the choices, and wait for the user to select one narrator:

bashCopy code
higgsfield voices list --json

Keep the selected voice's exact id and type (preset or element). Never invent or auto-pick a voice unless the user explicitly delegates it.

Generate one completed seed_audio job per narration block, always with the same voice:

bashCopy code
higgsfield generate create seed_audio \
  --prompt "<Block N narration only>" \
  --voice_type "<preset|element>" \
  --voice_id "<voice UUID>" \
  --wait \
  --json

Record every audio job UUID in block order. Regenerate only a failed or excessively long take. Shorten that block or adjust --speech_rate modestly when needed. Do not begin Phase 5 until all N audio jobs are complete.

Phase 5 — generate every clip second

Generate one completed 10-second gemini_omni clip per block. Attach the same style key to every call:

bashCopy code
higgsfield generate create gemini_omni \
  --prompt "<Block N video prompt>" \
  --image "<STYLE_KEY_ID>" \
  --duration 10 \
  --resolution 720p \
  --aspect_ratio 16:9 \
  --wait \
  --json

Use 9:16 when selected. Record every video job UUID in block order. Independent jobs may run concurrently inside this phase, but the audio-phase barrier is strict. Re-submit only failed blocks. Never silently replace gemini_omni; inspect the live video catalog if the model is unavailable.

Phase 6 — assemble immediately

Create blocks.json with at least two ordered block pairs. The CLI model contract requires typed references, so use the generic completed-job types:

jsonCopy code
[
  {
    "video": {"id": "<clip 1 job UUID>", "type": "video_job"},
    "audio": {"id": "<voice 1 job UUID>", "type": "audio_job"}
  },
  {
    "video": {"id": "<clip 2 job UUID>", "type": "video_job"},
    "audio": {"id": "<voice 2 job UUID>", "type": "audio_job"}
  }
]

Submit the server-side assembler immediately:

bashCopy code
higgsfield generate create explainer_video \
  --items @blocks.json \
  --width 1280 \
  --height 720 \
  --wait \
  --json

Use --width 720 --height 1280 for vertical. When subtitles are enabled, add the chosen font:

bashCopy code
--subtitles '{"font":"patrick"}'

The assembler keeps each block at exactly 10 seconds: it centers short voice takes, pitch-safely speeds small overruns, never stretches video, concatenates blocks in order, and optionally burns timed captions. Total duration is exactly N × 10 seconds.

Do not use local ffmpeg, the legacy assembly scripts, or the monolithic video_explainer job.

Checkpoints and recovery

• Before Phase 5: require one style key, exactly N narration lines and prompts, one selected voice, and N completed audio jobs.
• Before Phase 6: require N completed video jobs and exact one-to-one block pairing with no missing or duplicate IDs.
• Preset missing: refresh preset list; never reuse or fabricate an ID.
• Preset resolve failure: verify workspace selection and retry once.
• Style drift or realism: strengthen the shared STYLE and NEGATIVE text, then regenerate only that clip.
• Timeout: rejoin with higgsfield generate wait <job_id> --json; never duplicate a running job.
• Two identical failures mean the prompt or parameters must change.

Deliver

Return the final assembled video URL, exact duration, aspect, narration language, selected style, narrator, subtitle status, and a Sources list for researched topics. Keep intermediate job IDs and loose asset URLs internal unless requested.