You think X, actually Y (three-beat ladder)

Template (segment lengths assume duration 10 → 30% / 30% / 40%; rescale for 5 or 15): Beat 1 YOU THINK {X}; Beat 2 ACTUALLY: IT’S {Y}; Beat 3 BUT WHAT REALLY DECIDES IS {Z} — X / Y / Z come from the user’s brief (you compress each to ≤6 words for on-screen legibility). Hard cuts and kinetic energy between beats. English subtitles with exact windows. Default duration is 10 when unspecified.

Dependencies: WERYAI_API_KEY + Node.js 18+. scripts/video_gen.js + WERYAI_VIDEO_API.md. No other Cursor skills required.

Prerequisites

• - WERYAI_API_KEY must be set; Node 18+. Prefer public https image URLs; if the assembled scripts/video_gen.js supports local file paths, review/verify the script and explicitly consent before local read-and-upload to WeryAI.
• Non-empty model in every submit-* / wait JSON.
• Each successful task submission may consume credits; re-submit creates new paid tasks.

Security, secrets, and API hosts

• - WERYAI_API_KEY: Secret; never commit.
• API hosts fixed in video_gen.js.
• Local image handling disclosure: Prefer public https image URLs. If the assembled scripts/video_gen.js supports local file paths, it may read a local image and upload it to WeryAI to obtain a public URL; require review / verification and explicit consent before using that path.
• Higher assurance: Prefer a short-lived or isolated environment; review scripts/video_gen.js before production use. Verify whether the runtime can read local image files and upload them to WeryAI, and obtain explicit consent before using that path.

Pre-submit gate (mandatory)

⚠️ No paid submit without explicit user confirmation. Do not call submit-text, submit-image, submit-multi-image, or wait until the user has explicitly approved the parameter table (including full expanded prompt with X / Y / Z filled and timed caption windows). Never infer consent from silence or vague “continue.” Explicit means clear intent such as confirm / go / approved / yes, generate (or equivalent).

Parameter confirmation table (before any submit): model, duration, aspect_ratio, resolution (Seedance only), generate_audio, negative_prompt (Kling only), image / images (if used), and the entire expanded prompt (not a summary).

Prompt expansion (mandatory)

• - Duration: Default 10; 5 or 15 when allowed and requested.
• Timeline scaling: Keep 30% / 30% / 40% of total runtime for the three beats (adjust seconds in the written prompt); or another explicit split the user approves.

User must supply (or approve) X, Y, Z concepts—if they only give a topic, propose three beats in the confirmation table before submit.

Visuals: Three distinct looks per beat (e.g. office runner → chess flip → neon brain schematic); arrows reversing, color grade shift each beat; motion throughout.

Typography: Bold sans, outline, lower third; ALL CAPS optional for punch—keep lines short.

Negatives (Kling): illegible text, watermark.

Confirmation: The full prompt including filled X/Y/Z inside the timed block appears only in the pre-submit table; see ## Pre-submit gate (mandatory) — wait for explicit approval before submit-* or wait.

Niche checklist

• - Avoid demeaning groups; critique ideas, not people.
• If user’s Z is vague, default to SKILL / LUCK / COGNITION style abstractions they confirm.

Workflow

1. 1. Choose path: text-to-video, single image, or multi-image (SEEDANCE20 only, ≤3 URLs).
2. Extract or draft X, Y, Z; if the user only gave a topic, propose three beats and get their OK on concepts before expanding the full cinematic prompt.
3. Expand prompt (mandatory): Build the full production prompt (three distinct looks, motion, timed English captions per beat, negatives for Kling). Validate length against prompt_length_limit for the chosen model.
4. Pre-submit gate: Show the parameter confirmation table with the complete expanded prompt. Stop until the user explicitly confirms or requests edits.
5. Submit (async, default): After explicit confirmation, run submit-text, submit-image, or submit-multi-image (same JSON shape as wait; see WERYAI_VIDEO_API.md). 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 successful accept (e.g. taskId / batchId returned), immediately tell the user the id(s), short status (e.g. queued), and ask whether to continue with status polling or wait (block-until-done).
7. Continue (user-driven): Only after they agree, poll status --task-id <id> at reasonable intervals or run wait --json '…' with the same payload — per user choice.
8. Return results: [Video](url) when URLs exist; otherwise explain errorCode / errorMessage. Never wrap playable links in code fences.

CLI reference

After confirmation — async submit, then status (user opts in to polling):

~~~sh
node scripts/videogen.js submit-text --json '{"model":"KLINGV30PRO","prompt":"…","duration":10,"aspectratio":"9:16","generateaudio":true,"negative_prompt":"illegible text, watermark, garbled captions"}'
node scripts/videogen.js submit-image --json '{"model":"SEEDANCE20","prompt":"…","image":"https://…","duration":10,"aspectratio":"9:16","resolution":"720p","generate_audio":true}'
node scripts/videogen.js status --task-id ID>
~~~