cinematic-story-video-gen（电影故事视频生成）

Cinematic story & film-language video

Cinematic story video in one line: dramatic light, shallow depth of field, widescreen framing, believable skin, optional orange–teal contrast, subtle film grain, and camera grammar (dolly, track, orbit, foreground occlusion)—plus piano / strings / ambient-style audio when models allow.

Use when

• Someone asks for a cinematic, film look, movie-like, or mood / atmosphere clip with story or emotion (not only a generic filter).
• Brand film, hero moment, dialogue-free beat, or B-roll that should read like narrative cinema.
• Text-to-video or image-to-video (single reference still) with 16:9 (default), 9:16, or 1:1.
• Users mention orange and teal, film grain, bokeh, rack focus, slow push-in, or cinematic lighting.

Narrative cinema in motion: strong light–shadow hierarchy, shallow depth of field, widescreen framing, believable skin tones, warm key vs cool fill (orange–teal when it fits the scene), subtle film grain, and camera grammar that reads like a scene—not a filter dump.

Model choice (this package): ⭐ Best → KLING_V3_0_PRO (default for final emotion, skin, and lighting). 👍 Good → KLING_V3_0_STA (balanced cost). ⚡ Fast / draft → SEEDANCE_2_0 with resolution 720p (explicit resolution control; omit negative_prompt for this model). Kling rows must omit resolution from JSON.

Dependencies: WERYAI_API_KEY + Node.js 18+. This folder includes references/WERYAI_VIDEO_API.md (CLI/API reference); scripts/video_gen.js is supplied by your sync / publish step (same as other WeryAI video skills). Full commands and JSON fields: references/WERYAI_VIDEO_API.md. No other Cursor skills required.

Prerequisites

• WERYAI_API_KEY must be set before running video_gen.js.
• Node.js 18+; prefer public https image URLs. If the bundled runtime supports local file paths, review scripts/video_gen.js, verify read-and-upload behavior, and obtain explicit consent before using a local path.
• Model: video_gen.js requires a non-empty model in every submit-* / wait JSON—no script default. List the chosen model_key in the confirmation table.
• Each successful **submit-* / wait** may consume credits; re-submit creates new paid tasks.

Security, secrets, and API hosts

• WERYAI_API_KEY: Treat as a secret; never commit its value. OpenClaw metadata lists requires.env / primaryEnv so installers know it is mandatory at runtime.
• API hosts (fixed in video_gen.js): Video tasks use https://api.weryai.com; the models list uses https://api-growth-agent.weryai.com. Do not document or rely on host overrides—only WERYAI_API_KEY is read from the environment.
• Local image handling: Prefer https URLs. If the bundled runtime can read local files, it may upload them to WeryAI to obtain a public URL before image-to-video; require review, verification, and explicit consent before that path.
• Higher assurance: Run in an isolated environment; review scripts/video_gen.js before production use.

Pre-submit gate (mandatory)

⚠️ No paid submit without explicit user confirmation. Do not call submit-text, submit-image, or wait until the user has explicitly approved the parameter table below, including the full expanded prompt (entire text, not a summary). Never infer consent from silence. Explicit means confirm / go / approved / yes, generate (or equivalent).

Parameter confirmation table (show before submit): model, duration, aspect_ratio, resolution (only if the chosen model supports it—never for Kling V3), generate_audio, negative_prompt (only for models that support it), full expanded prompt.

Workflow

1. Confirm the request matches this skill (text-to-video and/or single image-to-video).
2. Collect the user's brief, optional image URL, and tier:

- Default / best / "final" → KLING_V3_0_PRO - Balanced / good → KLING_V3_0_STA - Draft / cheap / fast → SEEDANCE_2_0 (use resolution 720p or 480p; do not send negative_prompt) - User may name a model_key explicitly if it appears in the frozen table for the active channel—then ignore tier defaults for model.

1. Collect aspect: default 16:9 for widescreen story; use 9:16 or 1:1 when the user targets mobile or square feeds.
2. Expand prompt (mandatory): Unless the user supplied a finished long prompt (within prompt_length_limit) and asked not to rewrite, expand per ## Prompt expansion (mandatory). Do not submit only the user's minimal words.
3. Validate JSON fields against the frozen tables: Kling → omit resolution; Seedance → include resolution when using that model; negative_prompt only for Kling (optional but recommended with a short cinematic-safe line).
4. Pre-submit gate: Show the confirmation table with the full expanded prompt; stop until explicit approval or edits.
5. Submit (async, default): After approval, run submit-text or submit-image. Do not start a long blocking wait in the same turn unless the user already asked to block until the video is ready.
6. Immediate notify: On success with taskId / batchId (or documented id fields), immediately tell the user: accepted, id(s), short queue note, and ask whether to continue with status polling or wait. Do not run long invisible wait loops without this choice.
7. Continue (user-driven): Only after the user agrees, poll status --task-id … or run wait --json '…' with the same payload—per user preference.
8. When presenting playable URLs, use Markdown inline links only (e.g. [Video](https://…)). Do not wrap those links in code fences.

CLI reference

Single source of truth: All concrete video_gen.js usage—submit-text / submit-image / submit-multi-image, status, wait, JSON field rules, image URL requirements, stdout keys, and copy-paste examples—is documented in references/WERYAI_VIDEO_API.md (next to SKILL.md under references/).

Portable invocation (OpenClaw {baseDir}): run from the skill package root; in OpenClaw listings, {baseDir} resolves to that root.

node {baseDir}/scripts/video_gen.js submit-text --json '{"model":"KLING_V3_0_PRO","prompt":"<expanded English prompt>","duration":10,"aspect_ratio":"16:9","generate_audio":true}'
node {baseDir}/scripts/video_gen.js status --task-id "<TASK_ID>"

Blocking until done (only if the user asked to wait): node {baseDir}/scripts/video_gen.js wait --json '…' with the same JSON shape—see references/WERYAI_VIDEO_API.md.

This skill adds only model-specific filtering on top of that doc: build JSON that satisfies ## Model and API constraints here (e.g. omit resolution for Kling V3; include resolution for Seedance 2.0; do not send negative_prompt to Seedance). Default rhythm: after explicit confirmation, prefer **async submit-* then status; use wait only when the user asked to block until completion—see the Agent UX** section in the API file for polling guidance.

Prompt expansion (mandatory)

video_gen.js does not auto-expand. Before **submit-* or wait, turn short input into a full English production prompt aligned with cinematic story** grammar—not generic “cinematic” adjectives only.

When: Brief, vague, missing shot/light/audio detail, or no usable prompt (topic only)—still build a full prompt.

What to add (checklist):

• Look: cinematic, dramatic lighting, shallow depth of field, widescreen composition, subtle film grain, orange–teal or warm key vs cool fill, natural skin tone priority for people.
• Lens & depth: subtle anamorphic-style cues if appropriate; rack focus or bokeh where it serves emotion; foreground objects for parallax and occlusion transitions.
• Camera grammar (pick 1–2 per clip): slow push-in (dolly in); lateral track; gentle orbit; foreground silhouette passing—describe speed and motivation (hesitation, revelation, intimacy).
• Narrative: a clear emotional beat or story moment within duration.
• Negative space & composition: rule-of-thirds, lead room, eye-line continuity if two subjects.
• Audio (mandatory when generate_audio is true): labeled Audio: block—cinematic underscore (generic: piano, strings, subtle pulse or ambient electronic pads), room tone, soft foley; no copyrighted tracks or recognizable branded cues. If the user wants silent video: set generate_audio: false and omit audio from the prompt; state that in the confirmation table.

Style anchors (weave naturally): *cinematic, dramatic lighting, shallow depth of field, widescreen, film grain, orange teal contrast.*

Length: Stay within prompt_length_limit 2000 for the selected model in the frozen table; drop lower-priority adjectives before losing the core beat.

Confirmation: The pre-submit table must include the full expanded prompt.

### Example prompts below are richness targets only—always derive from the user's actual brief.

Definition of done

Done when the user gets at least one playable Markdown inline link whose label is Video and whose target is https://…, or a clear failure with next steps. Pre-submit: parameters and full expanded prompt were explicitly confirmed. After submit: user was notified with task id(s) and chose status vs blocking wait. Submitted JSON must match the frozen row for the chosen model_key (especially: never send resolution to Kling V3; never send negative_prompt to Seedance 2.0). generate_audio defaults to true when the model supports audio unless the user requested silent output.

Boundaries (out of scope)

• Do not send resolution on KLING_V3_0_PRO or KLING_V3_0_STA—it is unsupported and may cause parameter errors.
• Do not send negative_prompt for SEEDANCE_2_0 (not supported).
• Do not rely on paths or unofficial docs outside this package for CLI/API details; use only references/WERYAI_VIDEO_API.md.
• Do not embed the secret value of WERYAI_API_KEY in files.
• Do not wrap user-facing playable URLs in Markdown code fences.

Example prompts

• Make a 16:9 cinematic brand mood film—slow push-in, orange-teal grade, shallow DOF, soft strings
• Woman learns bad news by a rainy window—slow dolly in, teal shadows, warm lamp, piano and strings
• Two colleagues reconcile in a lobby—orbit, shallow DOF, foreground plant wipes frame, subtle score
• Turn this photo into a moody film scene—gentle push-in, film grain, orange key on skin, cool background, ambient bed (image URL in JSON)

Model and API constraints (frozen for this skill)

Derived from repository alignment with WeryAI model metadata (2026-03-21 snapshot). Re-run node scripts/video_gen.js models after platform upgrades and refresh this table if values change.

Text-to-video

Image-to-video (single image)

Recommended models

Tier keywords: unspecified / quality-first → Best; "balanced" / "good" → Good; "draft" / "cheap" / "fast" → Fast (SEEDANCE_2_0).

Default parameters

Scenario: Text-to-video cinematic story

1. Capture subject, emotion, and setting (or accept a minimal brief).
2. Expand per ## Prompt expansion (mandatory); pick model_key from tier rules.
3. Pre-submit gate → explicit user confirmation (full prompt, correct optional fields per model).
4. submit-text → notify → user chooses status vs wait.

Scenario: Image-to-video cinematic story

Before use: Prefer public https:// URLs. Local paths only with verified runtime support and explicit consent.

1. Plan motion (focus pull, parallax, occlusion) and respect likeness if requested.
2. Expand prompt; include image and model_key in the confirmation table.
3. After explicit confirm: submit-image (or wait if the user pre-requested blocking).

Loop seam (optional)

If the user asks for a seamless loop, append at the end of the expanded prompt: seamless loop, perfect loop, ends where it begins—no separate API flag.

Skill: cinematic-story-video-gen.