360 Secure SkillHub
360 安全技能中心

Phantra AI: Hooks And First Seconds

为 Phantra 短视频场景撰写 2 到 4 秒的开场钩子镜头:以抓眼画面开场,在窗口内安排转折或揭示,并按时间窗口改写首镜和后续节奏;同时提供配文起草与基于本人已发布数据的评分参考。技能强调把数据当证据而非结论,不含整场结构、封面图与平台交付规则。

分类:创作
zeropage-onryo
zeropage-onryo
热度 5.21
最近更新 2026-11-03 01:58:28

安装

发给 Agent
帮我安装这个技能:npx secure-skills-cli add zeropage-onryo/phantra-ai-hooks-and-first-seconds
兼容性
通用兼容
联网要求
需要联网
数据出网
必须出网
Metadata
name
hooks-and-first-seconds
description
Writes the first shot of a Phantra scene so it holds attention (an arresting image, a turn or a reveal in the first seconds) and puts it into the scene's first timed shot, reading `what_works` and `score_caption` as evidence rather than verdicts. Use when someone asks for a hook, says the opening is slow, wants the first seconds rewritten, or asks for a caption. NOT for the whole scene's shape (write-scene), the cover image (thumbnails) or the platform's shape and length (platform-delivery).

Hooks and the first seconds

A short is decided in its first two or three seconds. The first shot of a scene is rendered as its own clip, so the hook is written as (0-Ns), the first timed window, and nothing comes before it. Writing costs nothing.

What holds

  1. Open on the arresting image, mid-action. The first frame is already the situation: hands on the tin, the door already open. No establishing shot, no walk into frame, no title.
  2. A turn or a reveal inside the window. By the end of the shot the frame is not what it seemed, or a second thing has entered. The question it leaves the viewer with is why they stay for shot 2.
  3. Serious reactions. People react as people do: a freeze, a glance, a step back. Never a cartoon face, hands thrown up, a look to camera.
  4. One shot, one setup. 2 to 4 seconds, one camera position, one action that fits it. The cut to shot 2 is the first edit; land it on the turn.
  5. Sound is half the hook. Name the first diegetic sound in the Sound line (the lid's scrape, a hinge) and nothing else; no music, no voice.
  6. End the scene abruptly. The last shot stops on its strongest frame: no fade, no lingering, no resolution the viewer can predict.

Shapes that work: the reveal (the frame holds a thing that should not be there); the interruption (an ordinary action stopped by something entering); the wrong scale or the wrong place (an everyday object, an impossible setting). Not: a title card, a logo, text on screen (the avoid list forbids overlays), "wait for it", a slow push-in on nothing.

Writing it into the scene

Follow write-scene's five parts. In the beats:

  • The first window is short, (0-3s) or (0-4s), and starts at 0.
  • Image first, then the turn: "(0-3s) Close on Sam's hands lifting the lid of a dented biscuit tin; inside, a house key on a bed of crumbs. He stops."
  • Name the reaction physically, never by its emotion ("he freezes", not "he is shocked").
  • Shot 2 answers with a cut: "(3-8s) Hard cut to the hallway ..."; the later windows are reflowed so they still fill the total length.
  • The look stays in the style block; the hook carries action only.

Save with write_scene. To change the opening of a saved scene, call write_scene with its scene_id, the WHOLE prompt, and the same refs, seconds and roles again (a rewrite keeps nothing it is not handed; with no refs the scene becomes a draft with no photos). It replaces the text the scene had and detaches any clips already on it, so show the new version and get a yes first; to keep the original and its clips, duplicate_scene with the new prompt instead. Then quote, the person's yes, and approve with the quote's tokens (price-and-render).

Evidence first: what_works

Before rewriting a hook for someone who has posted before, call what_works. It reads their own studio: what they kept per model and camera preset, how many tries a keeper takes, and what they posted with a scene behind it, with its latest numbers. Use its lines as evidence for how THEY open (the presets they kept, the posted scenes that did well), never as a rule, and when enough_data is false say so and write from craft.

The caption: score_caption

A caption rides beside the hook on the platform. Draft it with the person from THEIR claims: never invent a number, a quote, a review or an award. Then score_caption with the text: a score from 0 to 1 against the words, topics and hooks of their OWN posted videos that beat their median, with the evidence behind each movement. Read it as evidence, not a verdict: a low score against a handful of posts is a hint, and a neutral score (fewer than four posted videos with numbers) means no evidence, which it says. Show the evidence lines, offer one revision, and let the person choose. Never loop on score_caption until a caption reads high.

A worked example

Person: "The opening of scene 403 is slow; make it grab."

  1. scene 403: shot 1 is "(0-5s) Sam walks into the kitchen and looks around." Nothing happens in it.
  2. what_works: enough_data true; the preset they kept most was handheld_rough, and posted names the scene with the best numbers; scene on that scene_id shows it opened on a reveal.
  3. Rewrite shot 1 as "(0-3s) Close on Sam's hands lifting the lid of a dented biscuit tin; inside, a house key on a bed of crumbs. He stops." and reflow the windows to fill 10 s.
  4. Show the whole scene; on a yes, write_scene scene_id=403 with it and the same refs and seconds as before.
  5. Caption drafted from what they told you; score_caption; show the evidence lines and the one revision they suggest.

Do not

  • Never put text, a logo or a voice-over pitch in the first shot.
  • Never invent numbers, quotes, reviews or results for the caption.
  • Never say a hook "will perform"; say what the person's own evidence shows.