360 Secure SkillHub
360 安全技能中心

Multi Variant Scripting

核心功能是基于单份产品brief,生成2-4个适配不同受众(技术精英/普通用户/开发者)、风格(电影感/导览式/讲解式)的差异化视频脚本。采用场景交错写作法避免同质化、术语渗透与篇幅失控,配套字数目标与交付校验清单。典型场景为需制作多版本产品演示视频的需求,输入为产品brief,输出为含分镜、纯旁白、素材要求的规范脚本文件。

nissan
nissan
热度 41.77
最近更新 2026-05-11 17:37:34
分类:创作依赖:纯提示词应用场景:文案创作 / 故事创作 / 模板

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add multi-variant-scripting
Metadata
name
multi-variant-scripting
version
1.0.0
description
Produce 2-4 genuinely distinct video script variants from a single brief — technical/cinematic (A), general/walkthrough (B), developer/explainer (C). Uses interleaved scene-by-scene writing to prevent drift, jargon bleed, and length creep. Includes word count targets and an output checklist for Finn handoff.
metadata
{"openclaw":{"emoji":"🎬","requires":{"bins":[],"env":[]},"primaryEnv":null,"network":{"outbound":false,"reason":"Document writing only. No network calls required."},"security_notes":"All operations are performed locally. No data leaves the user's machine."}}

Skill: Multi-Variant Scripting

Owner: Sara
Version: 1.0
First used: 2026-03-24 (Reddi Agent Protocol — 3-variant pipeline)


What This Does

Produces 2–4 distinct script variants for the same product from a single brief. Each variant serves a different audience or communication style. Variants must be genuinely distinct — not the same script with synonyms swapped.

When to use this skill: Any time the demo-video playbook calls for 2+ variants. Read this before writing the first word of any script.


The Variant Differentiation Test

Before writing a single scene, define these three things for each variant:

  1. What does this audience already know? (Don't explain what they know — assume it)
  2. What's the ONE thing this variant needs them to feel or understand? (Not a list — one thing)
  3. What's the register? (cinematic / conversational / technical)

If two variants have the same answers to all three → they are not distinct → collapse them into one.

Running this test upfront prevents the most common failure: writing three scripts that tell the same story three different ways, rather than three genuinely different stories about the same product.


Variant Labelling Convention

Use these labels consistently across all projects so Finn knows the production approach without reading the script:

Label Audience Style VO Density
A Technical or sophisticated Cinematic — minimal VO, UI carries story Sparse
B General / accessible Walkthrough — guided, second person Medium
C Developer / architect Explainer — structured, covers architecture Dense

Word Count Targets

Check these before calling any script done:

Target Length Narration Word Count
60s 110–130 words
90s 160–195 words
2min 220–260 words
3min 330–390 words

At natural TTS pace: ~130 words/min. Count words in the clean narration section only (no stage directions).


Interleaved Writing Technique

The wrong way: Write A fully → Write B fully → Write C fully

By the time you write C, it has unconsciously borrowed A's metaphors and B's sentence structure. C ends up sounding like a compressed remix of the first two.

The right way (interleaved): Write Scene 1 for A, B, C → Write Scene 2 for A, B, C → continue scene by scene

This forces explicit differentiation at each narrative beat. When you write Scene 2 for C, you've just written Scene 2 for A and B — so you're actively thinking "how is C different here?" rather than just continuing a flow.


Common Failure Modes

Drift

What it is: Writing C after A and B causes C to unconsciously borrow A's metaphors and B's structural patterns.
Fix: Use interleaved writing. Never fully complete one variant before starting the others.

Jargon Bleed

What it is: Technical terms written for C leak into B because they were written in the same session.
Fix: Before writing, define a jargon whitelist per variant. Terms on C's whitelist should not appear in B. Terms on B's whitelist should not appear in A unless they're genuinely plain-language.

Example:

  • A whitelist: (none — no jargon, let the UI speak)
  • B whitelist: "agent", "automates", "dashboard"
  • C whitelist: "MCP protocol", "orchestrator", "tool-call", "mesh routing"

Length Creep

What it is: All variants drift toward the same word count because the writer fills to a comfortable length.
Fix: Count words in the clean narration section after writing each scene. Stop when you hit the ceiling. Tighten, don't pad.


Output Checklist — Before Handoff to Finn

Run this check on all variants together, not one at a time:

  • Each variant has a clean narration section (narration only — no stage directions, no scene headers)
  • Word counts checked against length targets (count them, don't estimate by feel)
  • No variant shares an opening line with another (first sentence must be unique per variant)
  • Jargon that appears in C does not appear in B
  • Jargon that appears in B does not appear in A
  • Screenshot/recording requirements section is complete and specific
  • "Notes for Finn" section includes timing gotchas, transition notes, any unusual assembly requirements
  • Differentiation test re-run on final drafts — confirm answers to all three questions are still distinct

Handoff Format

Each script file must follow the standard format:

# Video Script [A/B/C] — [Title]
**Version:** 1.0
**Target length:** X seconds
**Audience:** [who]
**Style:** [description]

## Scene-by-scene
### Scene N — [Title] (M:SS–M:SS)
**Screen:** [page/state]
**Action:** [interaction if any]
**Narration:** "[exact spoken words]"
**Music/mood:** [description]

## Full narration (clean read)
[narration only — this is the direct TTS input]

## Screenshot/recording requirements
[specific assets needed, with reuse notes]

## Notes for Finn
[timing, transitions, assembly gotchas]

All three script files are handed to Finn together. Finn does not start production until Nissan has approved all scripts.