scenario-inspiration

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Scenario Inspiration

Scenario 灵感

Overview

概述

Asked for inspiration, an agent averages. It returns the most typical answer for the brief, so every user with a similar brief gets the same look. This skill trades that for four phases: prime, widen, choose, board. Connection and scope: see the

scenario

skill in this repo. If a sibling skill named here is missing from your available skills, ask the user to install it (

npx skills add scenario-labs/skills --skill <name>

); unattended, proceed from tool schemas and flag the gap.

Four moves carry most of the value. Name the reflex answer for this brief and the answer of someone merely avoiding that reflex, then rule out both, because an agent told to be original lands on the second-obvious answer. Randomize the operator, never the subject: what varies is the transformation applied to the brief, and a random subject is just noise. When you generate candidates yourself instead of searching for them, ask yourself for a distribution rather than an answer, and work from its tail. Then make the user pick, because options that never resolve are a gallery, not direction.

当被要求提供灵感时，Agent往往会给出平均水平的结果。它会返回符合需求说明的最典型答案，因此有类似需求的用户都会得到相同的视觉效果。本skill用四个阶段替代这种模式：启动锚定（prime）、拓宽发散（widen）、筛选抉择（choose）、整合成板（board）。关联与适用范围：参见本仓库中的

scenario

skill。如果此处提到的同级skill不在你的可用skill列表中，请让用户安装（

npx skills add scenario-labs/skills --skill <name>

）；若为无人值守模式，则基于工具schema继续执行并标记缺失项。

核心价值来自四个关键操作。先说出这个需求的本能反应答案，以及仅仅为了避开本能反应而给出的答案，然后将两者都排除——因为被告知要原创的Agent往往会落在第二个显而易见的答案上。要随机化操作方式，而非主题：变化的是应用于需求的转换方式，随机的主题只是噪音。当你自行生成候选方案而非搜索时，要着眼于分布而非单个答案，从分布的尾部入手。最后让用户来选择，因为永远没有定论的选项只是画廊陈列，而非方向指引。

Quick reference

快速参考

Phase	Do	Detail
1. Prime	Locked constraints, the open axis, both reflexes, what counts as a hit	references/widen.md
2. Widen	Four lanes plus wildcard draws; over-gather, near and far	widen.md, references/sources.md
3. Choose	Three or four mutually exclusive directions one named axis apart, plus an escape hatch	references/choose.md
4. Board	The pick becomes a collection, every reference annotated with its job	references/moodboard.md

Four lanes feed phase 2. Ask the user first, for three things they already love and one they cannot stand; unattended, take those from the task instructions and continue without them if it names none. Then

search

with

target: "assets"

twice, once over the team's own work (

public

omitted) and once over the public catalog (

public: true

), and last the open web, which is not a

search

call. Public asset hits carried

metadata.prompt

, the wording that produced them, and

target: "models"

hits carried

exampleAssetIds

at authoring time: free style research either way.

images: {like, unlike}

steers by example in both directions.

Draw far domains yourself with scripts/wildcard.py, run from the skill directory (

python3 scripts/wildcard.py --count 4

). A model asked for something random samples its own habits; the script samples a corpus and prints its seed, so a draw can be replayed or deliberately never repeated.

阶段	操作内容	详情说明
1. 启动锚定（Prime）	明确约束条件、开放维度、两种本能反应，以及合格标准	references/widen.md
2. 拓宽发散（Widen）	四条路径加随机抽取；广泛收集远近不同的参考素材	widen.md, references/sources.md
3. 筛选抉择（Choose）	三到四个沿同一维度分布、互斥的方向，外加一个备选出口	references/choose.md
4. 整合成板（Board）	将选中的方向整理为集合，为每个参考素材标注作用	references/moodboard.md

第二阶段由四条路径提供输入。首先询问用户，找出三个他们已经喜欢的和一个完全无法忍受的元素；若为无人值守模式，则从任务说明中提取这些信息，如果未提及则跳过。然后使用

target: "assets"

执行两次

search

：一次搜索团队自有作品（省略

public

参数），一次搜索公共目录（

public: true

），最后是开放网络搜索——这不属于

search

调用。公共资产结果带有

metadata.prompt

（生成该资产的提示词），而

target: "models"

的结果在创作时带有

exampleAssetIds

：两种方式都可用于免费风格调研。

images: {like, unlike}

可通过正反示例引导方向。

使用scripts/wildcard.py自行抽取跨领域参考，在skill目录下运行命令：

python3 scripts/wildcard.py --count 4

。要求模型生成随机内容时，它采样的是自身的训练习惯；而这个脚本采样的是语料库并输出种子值，因此抽取结果可以复现，也可以刻意避免重复。

Worked example: a puzzle game's world map, nothing decided

实操示例：解谜游戏的世界地图（尚无任何设定）

Prime. Locked: 16:9, readable at phone size. Open: everything else. Reflex: candy-colored isometric islands. Second reflex: the same islands, muted and "cozy". Both are out of bounds now. A hit is a map whose regions read apart in grayscale. Run the obvious query once and keep it as the baseline to beat.
Widen.
```
search
```
the project for anything already on brand, then
```
public: true
```
at
```
query_semantic_ratio: 0.8
```
and again at the keyword default, reading the prompts on the best hits. Draw four wildcards, run each as its own query, and keep near, middle, and far finds rather than only the strangest.
Choose. Name the axis first (here, how the world is depicted), then four positions on it. Each gets a title, an intent line, two or three references shown with
```
asset_display
```
, and who it wins for and when it fails. Say what you cut. Ask which is closest and what they would take from another, and offer the escape hatch. When nobody can answer, take the pick from the task instructions, else choose the option that best satisfies the hit line from prime, say which and why, mark it provisional, and keep going.
Board.
```
collection_create
```
the winner, then copy in every foreign reference: public assets belong to their own team, and
```
collection_add_assets
```
on one returned 403 at authoring time.
```
asset_download
```
,
```
curl -L
```
,
```
upload_asset
```
.
```
asset_update
```
each with its reason, tag the anti-references, hand the collection id to
```
scenario-image
```
or
```
scenario-consistency
```
.

启动锚定（Prime）。 约束条件：16:9比例，在手机尺寸下可读。开放维度：其他所有内容。本能反应：糖果色的等距岛屿。第二本能反应：同样的岛屿，但色调柔和、走“温馨”风格。这两种现在都被排除。合格标准：地图的各个区域在灰度下也能清晰区分。先运行一次最直观的查询，将结果作为需要超越的基线。
拓宽发散（Widen）。 在项目中
```
search
```
所有符合品牌调性的内容，然后用
```
public: true
```
分别以
```
query_semantic_ratio: 0.8
```
和默认关键词模式搜索，查看最佳命中结果的提示词。抽取四个随机（wildcard）关键词，每个单独作为查询运行，保留近、中、远相关的发现，而不是只留最奇特的。
筛选抉择（Choose）。 先明确维度（这里是“世界的呈现方式”），然后给出该维度上的四个位置。每个方向都要有标题、核心意图、用
```
asset_display
```
展示的两到三个参考素材，以及适用人群和不适用场景。说明你排除了哪些内容。询问用户哪个方向最接近他们的想法，以及他们想从其他方向中借鉴什么，同时提供兜底选项。如果无法得到用户反馈，则从任务说明中提取选择依据；否则选择最符合启动锚定阶段合格标准的选项，说明选择的内容和原因，标记为暂定，然后继续推进。
整合成板（Board）。 用
```
collection_create
```
创建选中方向的集合，然后复制所有外部参考素材：公共资产属于其所属团队，创作时调用
```
collection_add_assets
```
添加公共资产会返回403。因此需要用
```
asset_download
```
、
```
curl -L
```
下载后再
```
upload_asset
```
。用
```
asset_update
```
为每个资产标注作用，给反例参考打标签，然后将集合ID交给
```
scenario-image
```
或
```
scenario-consistency
```
使用。

Common mistakes

常见错误

Skipping prime and searching the brief's own words, which returns the category reflex sorted by relevance.
Four options that are one idea in four colors. If picking A does not rule out B, they are not directions.
Presenting only the survivors. Naming what was cut, and why, is what makes the rest credible.
Calling a set varied without checking it against what the literal query already returned.
Treating every strange find as a keeper. It earns a slot only when you can name what it connects to and what it is worth.
Boarding whatever came back. Every reference needs a job (light, color, composition, material, shape, subject, environment) or it is decoration.
Mixing lighting worlds on one board: a hard-flash reference and a soft-window reference cancel at generation time.
Handing a model an 18-image board. Reduce to three to six role-tagged references first.
Shipping the undirected default when the brief asked for none of it: subject centered and facing the lens, posed, evenly lit, everything in focus, a processed gloss. Every axis a prompt leaves open (framing, moment, lens, time of day, light) is filled with the model's most average answer, so name each one.

跳过启动锚定阶段，直接用需求的原词搜索，这会返回按相关性排序的品类本能反应结果。
四个选项本质上是同一个想法换了四种配色。如果选A并不意味着排除B，那它们就不是独立的方向。
只展示最终保留的选项。说明你排除了什么、为什么排除，才能让剩下的选项更有说服力。
没有和字面查询的结果做对比，就声称一组选项是多样化的。
把所有奇特的发现都当成有用的素材。只有当你能说出它和主题的关联以及它的价值时，它才有资格被保留。
把所有返回的结果都放进情绪板。每个参考素材都要有明确的作用（光影、色彩、构图、材质、形状、主题、环境），否则它只是装饰。
在同一个情绪板上混合不同的光影体系：硬闪光参考和软窗光参考会在生成时相互抵消。
给模型喂一张有18张图的情绪板。要先精简到3到6张带有角色标签的参考素材。
当需求没有要求时，使用无方向的默认设置：主体居中、面向镜头、摆拍姿势、光线均匀、全焦清晰、带有精修光泽。提示词中每一个留空的维度（构图、瞬间、镜头、一天中的时间、光线）都会被模型填充为最平均的答案，因此要明确指定每个维度。