图像生成
- 作者仓库星标 393
- 许可证 Apache-2.0
- 作者更新于 实时读取
- 作者仓库 gradient-bang
- 领域
- AI 智能
- 兼容 Agent
-
- Claude Code
- Cursor
- Cline
- Codex
- Windsurf
- Gemini CLI
- +20
- 信任分
- 94 / 100 · 已通过审计
- 作者 / 版本 / 许可
- @pipecat-ai · Apache-2.0
- Token 消耗评级
- 低消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部 API Key
- 需要 · OpenAI
- 兼容的系统
- 未声明(默认跨平台)
- 底层运行要求
- 无特殊要求
- 文件与系统权限
-
- 只读
- 允许写入 / 修改
- 读取环境变量
- 网络行为
- 允许外网请求
- 安装命令数
- 26 条
档案由构建时根据 SKILL.md 与安装命令自动衍生,可能与作者实际意图存在差异。
需要注意: 未限定 allowed-tools,默认拥有全部工具权限。
---
name: newspaper
description: Generate Gradient News & Observer assets (banners, front pages, prompt experiments) by dispatchi…
category: AI 智能
runtime: 无特殊运行时
---
# newspaper 输出预览
## PART A: 任务判断
- 适用问题:提示词、Agent 工作流、模型评估或自动化推理。
- 输入要求:目标材料、限制条件、期望输出和验收方式。
- 证据边界:围绕“Usage / All output goes under artifacts/ / Banner workflow”读取原文规则,不把推断写成作者承诺。
## PART B: 执行结果
- **01** 任务判断:确认你的需求是否属于提示词、Agent 工作流、模型评估或自动化推理,并标出输入、限制和预期结果。
- **02** 执行计划:优先按“Usage / All output goes under artifacts/ / Banner workflow”拆成步骤,说明每一步会读取什么、修改什么、产出什么。
- **03** 交付结果:给出可复制的命令、文件改动、检查清单或内容草稿,并说明如何继续迭代。
- **04** 风险边界:结合 读取文件、写入/修改文件、读取环境变量、会按任务需要访问外部网络、需要准备 OpenAI API Key 给出执行前确认项。
## Running Rules
- 读取文件、写入/修改文件、读取环境变量;会按任务需要访问外部网络;需要准备 OpenAI API Key。
- 先小样例验证,再放大到真实任务。
- 交付时同时给结果、检查口径和下一步迭代建议。 原文出现了 `/newspaper`、`/news-front-page` 这类斜杠命令;如果你的 Agent 支持命令触发,优先用命令开场,再补充目标和边界。
告诉 Agent 目标文件或材料、期望结果、不可改范围、是否允许联网或执行命令。本 Skill 的权限画像是:读取文件、写入/修改文件、读取环境变量。
先用一个小任务确认它会围绕“Usage / All output goes under artifacts/ / Banner workflow”工作;涉及文件或命令时,先看 diff、日志、预览或测试结果。
检查最终产物是否包含明确结果、必要证据和下一步动作;如果输出泛泛而谈,就补充输入、边界和验收标准后重跑。
---
name: newspaper
description: Generate Gradient News & Observer assets (banners, front pages, prompt experiments) by dispatchi…
category: AI 智能
source: pipecat-ai/gradient-bang
---
# newspaper
## 什么时候使用
- 把 AI / Agent方向的常用动作沉淀成 Agent 可调用的技能 适合处理AI Agent、提示词、模型评估与自动化推理,核心价值是把输入、判断、执行、验证和交付边界固定下来,避免 Agent 泛泛回答。 把任务拆成可执行、可检查…
- 面向提示词、Agent 工作流、模型评估或自动化推理,优先处理能明确输入、步骤和验收标准的工作。
## 需要提供什么
- 目标材料、目录范围、期望结果和不可改动内容。
- 是否允许联网、执行命令、读写文件或调用外部服务。
## 执行规则
- 围绕「Usage / All output goes under artifacts/ / Banner workflow」组织步骤,不把推断写成作者事实。
- 读取文件、写入/修改文件、读取环境变量;会按任务需要访问外部网络;需要准备 OpenAI API Key。
- 先跑小样例,确认结果可检查后再扩大任务范围。
## 输出要求
- 给出最终产物、关键证据、验证方式和下一步动作。
- 信息不足时标记 unknown,不编造命令、平台或依赖。 作者原文负责流程事实;仓库文件负责来源和命令;流狐只补充适用场景、限制和质量判断。
skill "newspaper" {
输入层 -> 用户目标 + 目标文件 + 禁止范围 + 验收标准
上下文层 -> Usage / All output goes under artifacts/ / Banner workflow
规则层 -> SKILL.md 触发条件 / 执行顺序 / 输出格式
运行层 -> 无特殊运行时 | 读取文件、写入/修改文件、读取环境变量 | 会按任务需要访问外部网络
安全层 -> 需要准备 OpenAI API Key + 小任务验证 + diff / 日志复核
输出层 -> 可复制结果 + 检查清单 + 下一步迭代
} Newspaper Asset Generator
Single entry point for all Gradient News & Observer image assets. Routes to the underlying script based on the asset type.
Usage
/newspaper <asset-type> [args]
Asset types:
banner— single wide masthead-style banner (e.g. CALL TO ARMS recruiting header). Handled below.front-page— full ten-section newspaper front page from real game events. Hands off to/news-front-page— invoke that skill directly instead.prompt-experiment— sweep prompt variants for front-page rendering. Callsnews-front-page-prompt-experiment(see src/gradientbang/newspaper/scripts/prompt_experiment.py for flags).
If the user passes an unrecognized asset type, list the three options above and stop — do not guess.
All output goes under artifacts/
artifacts/ is gitignored. Both copy markdown and rendered PNGs live there. Subdirectories per asset type:
artifacts/banners/artifacts/gradient-news-observer-regions/(front-page composited)artifacts/gradient-news-observer-images/(front-page one-shot)
Banner workflow
A banner is a single wide image with one dramatic headline + short body + call-to-action. Used for recruiting drives, special-edition broadsides, channel headers, etc.
Step 1 — gather copy from the user
Required inputs:
- Headline (largest type — e.g.
CALL TO ARMS) - Subhead (one line — e.g.
Pilots Wanted to Test the Combat Rework) - Body (1–3 short sentences setting the stakes)
- Call to action (e.g.
Join the test → #gradient-bang)
Optional:
- Kicker (small caps line above the headline; defaults to
THE GRADIENT NEWS & OBSERVER · SPECIAL EDITION) - Slug (short filename stem, kebab-case; defaults to a slugified headline)
If the user's prompt only gives you a topic ("call to arms for combat rework"), draft kicker/headline/subhead/body/CTA yourself in the newspaper's voice and confirm with the user before rendering — image-edit calls take ~2 minutes and cost real money.
Step 2 — write the copy markdown
Write to artifacts/banners/<slug>.md using this exact structure (each line on its own line, blank lines between blocks). Every visible word in the file is rendered verbatim — no extra hidden formatting.
<KICKER>
<HEADLINE>
<Subhead>
<Body sentences.>
<Call to action>
Keep total visible word count under ~50 — banners get cramped fast. If body runs long, trim it.
Step 3 — render
set -a && source .env.bot && set +a
uv run news-banner --copy-file artifacts/banners/<slug>.md
- Sources
.env.botbecause that's whereOPENAI_API_KEYlives in this repo (the script also auto-loads it from.envif present). - Defaults to
2048x1024(clean 2:1 banner). For wider Discord-banner shapes, append--size 2880x1024. Both dims must be multiples of 16; max edge 3840; max ratio 3:1. - Defaults to the masthead reference. For a busier full-page-style aesthetic, append
--reference src/gradientbang/newspaper/assets/references/full-page-style-fri-sat-pt.png. - Render takes ~120–150s. Stream stdout — the script prints the final PNG path on success.
- Pass
--forceto overwrite an existing PNG with the same name.
Step 4 — show the result
Read the generated PNG so the image is visible in chat, then summarize with the artifact path. The user reviews and either accepts, edits the markdown and re-renders, or asks for a fresh take (image-edit isn't deterministic — same input can produce a different layout).
Front-page workflow
Don't replicate the front-page logic here. Tell the user:
Front pages are a different animal — they pull real game events into a structured digest, then render a 2160×3840 ten-story page. Use
/news-front-page [duration](e.g./news-front-page 24h).
If they explicitly want to re-render an existing markdown front page without regenerating the digest, they can call news-front-page-image --front-page-md <path> directly.
Prompt-experiment workflow
This is a power-user tool for sweeping prompt variants. Run with --help first to surface the flags, then construct the right command for the user's question:
uv run news-front-page-prompt-experiment --help
Rules
- Always confirm copy before spending an image-edit call. Each one costs and takes ~2 min. Cheap to iterate on markdown, expensive to iterate on images.
- Only edit
artifacts/for copy and outputs. Don't write undersrc/for one-off banners. - Don't invent new script entry points. If the user wants something the existing scripts don't support (e.g. a different aspect or new layout), discuss the trade-off before adding code.
- Don't commit anything. The user reviews artifacts and decides what to share.
先判断是否适合
作者设计意图
作者的方法与取舍
边界和复核