AGENT SKILL

text-to-video-direct

Single-step text-to-video generation (Sora, Veo, Kling text mode, Hailuo). Use when you need quick exploratory shots or when motion is more important than precise framing.

License 许可
MIT
最近更新
2026-11-10
Stars
★ 0
Forks
0

01 前言 Front Matter

name
text-to-video-direct
description
Single-step text-to-video generation (Sora, Veo, Kling text mode, Hailuo). Use when you need quick exploratory shots or when motion is more important than precise framing.

03 SKILL.md 正文

Direct Text-to-Video Pipeline

Single model call from a text prompt to a clip. No intermediate still.

When to use

  • Exploratory or B-roll shots where you don't need to lock the opening frame.
  • Motion-heavy shots where the model's interpretation of motion matters more than composition.
  • Faster iteration when image-then-animate is overkill.

When NOT to use

  • Character-specific shots needing visual consistency — direct text-to-video drifts heavily across takes. Use text-to-image-to-video instead.
  • Talking heads — use voice-to-lip-sync over a generated character clip.

Inputs

  • Shot brief from scripts/storyboards/NN-*.md.
  • Text-to-video model from brief/tools-and-models.md.

Steps

  1. Compose prompt. Combine visual seed + motion description + style/vibe. Show to user.
  2. Generate. Call the configured text-to-video model. Run it through the provider that hosts the model: fal.ai and WaveSpeed models via the provider-runners skill (pass --output with the target folder); Replicate and MiniMax via their MCP servers. Save to generation/text-to-video/NN-shortname-vN.mp4. Save prompt + model + parameters to generation/prompts/NN-shortname-vN.md. Log it.
  3. Surface. Show clip. Offer retry, accept, or switch to text-to-image-to-video for tighter control.
  4. Accept. Copy to clips/raw/ and suggest /promote-take.

Common gotchas

  • Cost per retry is high. Text-to-video models are typically the most expensive call in any pipeline. Iterate the prompt carefully before each call — don't just hit retry.
  • Aspect / duration locked at gen time. Most models don't let you crop or extend after the fact cleanly. Set both correctly per the brief.

04 同仓库更多 Skills