kinetic-sand-video（动力沙视频）

Kinetic sand shaping video generation

Imagine: a blade eases into a bright yellow kinetic sand block—the cross-section opens from the center, grain friction and a perfect cut. Soft yet orderly satisfaction is what this skill is for.

Core beats:

• Clean slice: Sharp kerf, even color in the face, grain texture readable
• Mold release: Castle / star / geometry lifts out, edges crisp
• Layered collapse: Colored strata slip at the edge like a soft solid in slow motion
• Streaming pour: Sand falls evenly; pile-up feels fluffy and fine

Dependencies: scripts/video_gen.js in this directory + WERYAI_API_KEY + Node.js 18+. No dependency on other Cursor skills.

Prerequisites

• WERYAI_API_KEY must be set in the environment before running video_gen.js.
• Node.js 18+ is required. Image inputs must be public https URLs (no local file paths).
• Each successful wait run consumes WeryAI credits; re-running creates new paid tasks.

Security, secrets, and API hosts

• WERYAI_API_KEY: Treat as a secret. Only configure it if you trust this skill's source; it is listed in OpenClaw metadata as requires.env / primaryEnv so installers know it is mandatory at runtime (never commit it inside the skill package).
• API hosts (fixed in video_gen.js): Video tasks use https://api.weryai.com; the models list uses https://api-growth-agent.weryai.com. Only WERYAI_API_KEY is read from the environment—do not rely on URL-related environment variables.
• Higher assurance: Run generation in a short-lived or isolated environment (separate account or container), and review scripts/video_gen.js (HTTPS submit + poll loop) before production use.

Prompt expansion (mandatory)

video_gen.js does not expand prompts. Before every wait --json, turn the user's short or vague brief into a full English production prompt.

When: The user gives only keywords, one line, or loose intent—or asks for richer video language. Exception: They paste a finished long prompt within the model's prompt_length_limit and ask you not to rewrite; still show the full text in the confirmation table.

Always add (video language): shot scale and angle; camera move or lock-off; light quality and motivation; subject action paced to duration; one clear payoff for this niche; state 9:16 vertical when this skill defaults to vertical.

Length: Obey prompt_length_limit for the chosen model_key when this doc lists it; trim filler adjectives before removing core action, lens, or light clauses.

Confirmation: The pre-submit table must include the full expanded prompt (never a one-line summary). Wait for confirm or edits.

Niche checklist

• ASMR slice: blade entry, clean plane, lattice cross-section, slow pull-apart or restack.
• Texture: saturated layers, crisp edges, fine grain; macro top-down or 45°.
• Payoff: satisfying separation or collapse moment.

### Example prompts at the top of this file are short triggers only—always expand from the user's actual request.

Workflow

1. Confirm the user request matches this skill's scenario (text-to-video and/or image-to-video as documented).
2. Collect the user's brief, optional image URL(s), tier (best / good / fast) or an explicit model key.
3. Expand prompt (mandatory): Unless the user supplied a finished long prompt and explicitly asked not to rewrite it, expand the brief into a full English production prompt using ## Prompt expansion (mandatory) below. Do not call the API with only the user's minimal words.
4. Check the expanded prompt against the selected model's prompt_length_limit in the frozen tables in this document (when present); shorten if needed.
5. Verify duration, aspect_ratio, resolution, generate_audio, negative_prompt, and other fields against the frozen tables and API notes in this SKILL.md.
6. Show the pre-submit parameter table including the full expanded prompt; wait for confirm or edits.
7. After confirmation, run node {baseDir}/scripts/video_gen.js wait --json '...' with the expanded prompt.
8. Parse stdout JSON and return video URLs; on failure, surface errorCode / errorMessage and suggest parameter fixes.

CLI reference

node {baseDir}/scripts/video_gen.js wait --json '{"model":"…","prompt":"…","duration":5,"aspect_ratio":"9:16"}'
node {baseDir}/scripts/video_gen.js wait --json '…' --dry-run
node {baseDir}/scripts/video_gen.js status --task-id <id>

Definition of done

Done when the user receives at least one playable video URL from the API response, or a clear failure explanation with next steps. All parameters used must fall within the selected model's allowed sets in this document. The submitted prompt must be the expanded production prompt unless the user explicitly supplied a finished long prompt and asked not to rewrite it.

Boundaries (out of scope)

• Does not review platform compliance, copyright, or portrait rights; does not guarantee commercial usability of outputs.
• Does not provide non-WeryAI offline rendering, traditional edit timelines, or API field combinations not documented here.
• Does not hard-code absolute paths in the skill doc; {baseDir} means the skill package root (same level as SKILL.md).

Example prompts

• Rainbow kinetic sand one clean slice, neat cross-section, dense grains, vertical satisfying
• This image is a sandcastle: mold lifts off and the whole shape stands, motion
• Space sand collapse slow motion, colored layers like cake strata
• Kinetic sand cutting ASMR 9:16, clean slice cross-section

Default parameters

API validity (default KLING_V3_0_PRO): Text-to-video duration only 5 / 10 / 15; aspect_ratio only 9:16, 1:1, 16:9. Image-to-video aspect_ratio only 9:16, 16:9, 1:1. No resolution field—do not send it. For fast tier with VEO: text-to-video VEO_3_1_FAST, image-to-video CHATBOT_VEO_3_1_FAST, with duration fixed at 8, aspect_ratio only 9:16 or 16:9. When switching model_key, follow the allowed sets in this section’s model/API constraints and the API validity note above; do not send resolution to models that do not support it.

Text-to-video

User names colors / layers and action (slice / demold / collapse / pour)—skip asset hunting.

Tell me in one line: sand color or layer count, and the action.

Generation flow:

1. Collect palette + action; ask if missing.
2. Pick the best angle per action—slice: expose the cross-section; demold: whole shape clearing the mold; collapse: edge slippage; pour: top-down stacking.
3. If the user didn’t specify params, show this run and wait:

> Generating with the parameters below—reply confirm or say what to change: > - model: KLING_V3_0_PRO > - aspect_ratio: 9:16 > - duration: 5s > - generate_audio: true > - Loop seam: off by default (reply loop to append seamless loop, perfectly looping video, first and last frame identical to the prompt)

1. After confirmation (and optional loop keywords), run node {baseDir}/scripts/video_gen.js wait --json '…' (text-to-video); parse stdout.
2. Return URLs; note swaps (palette / action / loop).

Parameter configuration:

Expanded prompt: Compose at generation time per ## Prompt expansion (mandatory) from the user's actual brief—do not reuse fixed sample paragraphs.

Sand-shape image to motion

Public HTTPS photo of a kinetic sand shape → slice, press, or collapse starting from that look. Good for remixing assets or branded shapes.

Generation flow:

1. Confirm URL starts with https://.
2. Read shape (block / castle / free stack / mold imprint) and key colors; pick motion.
3. If unspecified, show and wait:

> Generating with the parameters below—reply confirm or say what to change: > - model: KLING_V3_0_PRO > - aspect_ratio: 9:16 > - duration: 5s > - generate_audio: true > - Loop seam: off (reply loop to enable)

1. Run node {baseDir}/scripts/video_gen.js wait --json '…' with image; parse stdout.
2. Return URLs.

Parameter configuration:

Expanded prompt: Compose at generation time per ## Prompt expansion (mandatory) from the user's actual brief—do not reuse fixed sample paragraphs.

Tips

Prompt boosts

• Cross-section: powdery matte cross-section, sharp color strata exposed, grain texture visible at the cut face
• Flow: individual grains tumble in slow motion, sand cascades like liquid but holds shape, fine particles catch the light mid-air
• Demold: mold lifts cleanly revealing intact geometry, crisp edges hold for one suspended second before softening, air pocket release at mold separation
• Soft solid: kinetic sand deforms like a soft solid, no bounce, zero rebound, pure absorption

Notes

• More saturation and more layers strengthen the slice payoff—aim for at least three colors.
• For loop playback, say so before generate; loop phrases are appended to the prompt.
• Clearer, more centered reference images track the shape better in image-to-video.

Note: Image URLs must be public HTTPS; private hosts or local paths will error at the API.