HiAPI Video Prompt Generator#

Before using this skill, run the update check once per session (it skips silently when offline; set HIAPI_SKIP_UPDATE_CHECK=1 to disable):

bash

node scripts/check-update.mjs

If it reports a required update, run the printed install command before continuing. If it reports an available update, tell the user and continue.

Use this skill when the user has a video idea, a product link, a research topic, or a brief, and needs a strong, executable video prompt before calling a video model.

This skill does not generate video. It produces a structured prompt that the user can paste into:

• hiapi-seedance-2-0-video-skill — high-quality text-to-video and image-to-video.
• hiapi-happyhorse-1-0-video-skill — lightweight text-to-video drafts.

The output is built so each fact, command, and UI label survives the trip to the model.

When To Use This Skill#

Use this skill when any of these are true:

• The user describes a video they want, but the brief is one or two sentences.
• The user gives a URL, GitHub repo, or document and asks for "a video about this".
• The user wants a product demo, teaching short, social short, explainer, or pitch.
• The user has a draft prompt that is vague, generic, or missing screen text and motion.

Skip this skill and route directly to the model skill when the user already has a final prompt and only wants to generate.

Defaults#

• Duration: 5 seconds. Never default to 30 — HiAPI's video models do not run that length.
• Aspect: 16:9 for horizontal demos and explainers, 9:16 for social shorts and vertical product teasers.
• Scene count: at micro-cut granularity (Seedance 2.0) — 4 blocks at 5 s, 5 at 8 s, 6 at 10 s, and up to 8 blocks at 15 s. At macro-beat granularity (HappyHorse 1.0) — 2 beats at 3 s, 3 beats at 5 s, 4 beats at 8 s, 5 beats at 10 s, 6 beats at 15 s. Either way the Output Contract still returns one block per scene.
• Resolution: Seedance 2.0 defaults to 720p, downgrades to 480p for fast drafts. HappyHorse 1.0 defaults to 1080p, downgrades to 720p when speed or cost matters; do not output 480p for HappyHorse.
• Language: Match the user's language. Keep product names, commands, and UI labels in their original casing.

The supported parameter sets are model-specific. This skill must pick a target first, then constrain duration and aspect to that target's list.

Seedance 2.0 ( hiapi-seedance-2-0-video)#

• Durations (seconds): any integer from 4 to 15.
• Resolutions: 480p, 720p, 1080p.
• Ratios (flag --ratio): 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, adaptive.
• Media modes: text-to-video; first-frame image-to-video; first+last-frame image-to-video; multimodal references.
• Mutual exclusion: do not mix first/last-frame fields with reference_image_urls, reference_video_urls, or reference_audio_urls.
• Reference limits: reference images plus first/last frames <= 9 images; reference videos <= 3, each 2-15 s and total <= 15 s; reference audio <= 3, each 2-15 s and total <= 15 s.

HappyHorse 1.0 ( hiapi-happyhorse-1-0-video)#

• Durations (seconds): any integer from 3 to 15.
• Resolutions: 720p, 1080p.
• Sizes (flag --size, not --ratio): 16:9, 9:16, 1:1, 4:3, 3:4. No 21:9.
• Seed: optional integer 0 to 2147483647; include only when the user asks for reproducibility.
• Input: text-to-video only. No image input.

Workflow#

1. Read the request. Classify it: product demo, teaching short, social short, explainer, pitch, historical, or visual concept.

2. Pick the target model. Default to hiapi-seedance-2-0-video for image-to-video, cinematic feel, or anything over 5 seconds. Use hiapi-happyhorse-1-0-video for a fast text-to-video draft.

3. Gather sources. Use materials the user supplied first. If they are insufficient, research:

   • Official product site, docs, GitHub repo, release notes.
   • Primary sources for the topic (papers, organization sites, official archives).
   • Reputable English-language background coverage if needed.
   • Avoid sources the user did not ask for if they would dilute or contradict the brief.

4. Extract visualizable facts. See references/source-extraction.md. Pull names, commands, UI labels, numbers, differentiators, visual hooks, and what the product should never claim.

5. Pick a pattern. See references/prompt-patterns.md. Adapt the pattern to the chosen duration and aspect ratio.

6. Write the prompt following the Output Contract below. Each fact, command, and UI label must show up either on screen, in narration, or in a labeled visual cue.

7. Verify the constraints. Confirm the duration is in the chosen target's allowed list (Seedance: any integer from 4 to 15; HappyHorse: any integer from 3 to 15). Confirm the aspect value matches the right flag for that target ( --ratio for Seedance, --size for HappyHorse) and is in its allowed list. For Seedance, choose exactly one media mode: first-frame, first+last-frame, or multimodal references. Confirm reference image/video/audio limits before writing a handoff command. Confirm screen text is short enough to render. Confirm nothing in the prompt fabricates a feature.

8. Hand off. See references/hiapi-handoff.md. Show the user the exact node scripts/... command for the chosen HiAPI skill so they can run it next.

Output Contract#

Return all of the following sections, in this order, in the user's language. Keep section headers in the user's language too.

1. Video Type — one of: product demo, teaching short, social short, explainer, pitch, historical, visual concept.

2. Target Model — hiapi-seedance-2-0-video or hiapi-happyhorse-1-0-video, with one reason.

3. Duration And Aspect — pick from the chosen target's allowed list above. Seedance: any integer from 4 to 15 seconds and one of 16:9|9:16|1:1|4:3|3:4|21:9|adaptive. HappyHorse: any integer from 3 to 15 seconds and one of 16:9|9:16|1:1|4:3|3:4. State which flag the target uses ( --ratio for Seedance, --size for HappyHorse).

4. Core Objective — one sentence: what the viewer should remember at the end.

5. Source Extraction Summary — five to ten bullet facts, each tagged [source] if pulled from a real source, or [creative assumption] if invented for staging. Never tag a guess as a source.

6. Narrative Through-Line — one sentence: the story arc from frame one to the last frame.

7. Scene Prompts — one block per scene. Each block has:

   • Time (e.g. 0.0s–1.2s)
   • Visual — what is on screen
   • On-Screen Text — exact strings, in quotes, short enough for the duration
   • Action / Camera — one dominant motion and one secondary ambient motion
   • Narration — one short line, or (none)
   • Transition — to the next scene

8. Required Screen Text — a deduplicated list of every quoted on-screen string. Preserve command syntax and tag text exactly.

9. Motion And Camera Control — list the dominant motions (typing, click, zoom, pan, dolly, reveal). Avoid particle effects unless the brief is artistic.

10. Style Requirements — color, lighting, typography, density, mood. One paragraph.

11. Negative Constraints — what the model must not show. Always include: no fabricated features, no logos the brief did not authorize, no on-screen text outside Required Screen Text.

12. Final Copy-Ready Prompt — a compact block suitable for --prompt. It MUST keep the scene order with short time cues (e.g. [0–1.2s]), every Required Screen Text string verbatim, the dominant motions, and at least one Negative Constraint. A flat description that drops these cues is wrong.

13. Handoff Command — the exact node scripts/... command for the chosen HiAPI skill, with duration, resolution, media mode, and the right ratio flag ( --ratio for Seedance, --size for HappyHorse) pre-filled. Prefix with the working directory the user must cd into first (the installed target skill directory).

Quality Bar#

A strong output is directed, not described. It tells the model what appears, what moves, what is typed, what is heard, what is grounded in the source, and what must not appear. If a fact in the Source Extraction Summary does not show up either on screen, in narration, or as a labeled visual cue, drop the fact or add a scene.

Do not fabricate facts. Product names, metrics, commands, UI labels, version numbers, testimonials, and feature claims must come from the brief or a cited source. Only staging choices — camera, layout, lighting, generic visual treatment, transition copy — may be tagged [creative assumption] when no source pins them. A missing fact is a question for the user, not a license to invent.

Routing And Adjacent Skills#

• For a still image starting frame before image-to-video, route to awesome-gpt-image-2-prompts for image prompts, then hiapi-gpt-image-2-skill to render it.
• For a library of proven video prompts to anchor on, see awesome-seedance-2-0-prompts.
• For broader model discovery, see hiapi-skills or the Remote MCP at https://mcp.hiapi.ai/mcp.

Examples#

text

$hiapi-video-prompt-generator Turn https://github.com/HiAPIAI/hiapi-skills into a 5-second product intro.

text

$hiapi-video-prompt-generator Make a 9:16 social short about why agent skills beat one-shot prompts.

text

$hiapi-video-prompt-generator Plan an 8-second image-to-video starting from outputs/product.png, soft camera, studio lighting.

Important Links#

• HiAPI Seedance 2.0 video skill: https://github.com/HiAPIAI/hiapi-seedance-2-0-video-skill
• HiAPI HappyHorse 1.0 video skill: https://github.com/HiAPIAI/hiapi-happyhorse-1-0-video-skill
• Get API key: https://www.hiapi.ai/en/register
• Pricing: https://www.hiapi.ai/en/pricing
• HiAPI docs: https://docs.hiapi.ai