🧩 dsh-claude-code-templates

中文 | English

中文

884 个 Claude Code 技能模板整包移植到 DeepSeek Harness（dsh）——不挑不筛全部打包，装完即用。

这是 davila7/claude-code-templates（MIT，30k+ stars）的 dsh 移植 （port）。来自上游的 884 个 SKILL.md（上游共 897 个，其中 13 个 Anthropic source-available 已移除，见下）与全部附属资源 （references/、examples/、assets/ 等 4160 个文件）逐字节原样复制， 技能内容一律未翻译、未改写，只加了一层最小的 dsh 载入外壳。

已移除 13 个 Anthropic source-available 目录。 上游 skills/ANTHROPIC_ATTRIBUTION.md 把这批文档技能标为 source-available（参考用途、不可再分发），其 LICENSE.txt 明文禁止 Reproduce / Distribute / sublicense / transfer 给任何第三方；上游公开托管 ≠ 授权我们散布，每个散布者独立负责。因此 document-processing/{docx, docx-official, pdf, pdf-anthropic, pdf-official, pptx, pptx-official, xlsx, xlsx-official} 与 scientific/document-skills/{docx, pdf, pptx, xlsx} 已从本 repo 移除，Apache-2.0 / MIT 内容全部保留。 需要这些技能请直接去上游取用：anthropics/skills 或 davila7/claude-code-templates。

✨ 功能特性

• 884 个技能全部打包：dsh 的 skill catalog 展示 860 个唯一技能（21 个名称跨类别重复、共 45 个文件，其中 14 组内容完全相同、7 组内容不同；catalog 以名称为键只展示一份，全部文件都保留在 repo）
• 29 个分类：ai-research、development、business-marketing、creative-design、scientific、security、web-development、workflow-automation 等，含 Anthropic Apache-2.0 官方技能（授权说明见上游 ANTHROPIC_ATTRIBUTION.md，原样保留）
• 零改动保真：THIRD_PARTY_NOTICES.sha256 钉住 5044 个文件的 SHA-256，npm prepare / npm prepack 自动校验，漂移即报错
• 含 Claude 官方插件技能：.claude-plugin/skills/owasp-security（OWASP 六标准安全参考，含 examples/）
• 模型可用：技能经 dsh 标准 skill 工具加载，无需额外配置

📸 效果

安装后 headless 会话的 skill catalog（真实运行输出，节选）：

- `2d-games`: 2D game development principles. Sprites, tilemaps, physics, camera.
- `3d-games`: 3D game development principles. Rendering, shaders, physics, cameras.
- `ab-test-setup`: When the user wants to plan, design, or implement an A/B test or experiment...
- `accessibility`: Audit and improve web accessibility following WCAG 2.1 guidelines...
- `git-commit-helper`: Generate descriptive commit messages by analyzing git diffs...

模型可以直接加载技能并执行其指令：

$ dsh --profile <P> "Load the skill named git-commit-helper and quote its first heading."
# Git Commit Helper

📦 安装

repo 当前为 private（开源与否待人工复核），可本地路径安装；公开后可用 GitHub 安装：

# 方式一：本地路径（当前 repo 未发布到 npm，也未开源）
dsh plugin --profile <P> add /path/to/dsh-claude-code-templates

# 方式二：GitHub（开源后可用；首次安装需允许构建脚本，见 dsh plugin 提示）
dsh plugin --profile <P> add github:GongYuanCaiJi/dsh-claude-code-templates

注意：新 profile 需先加入 headless 模式包（dsh plugin --profile <P> add @deepseek-ai/dsh-headless， 并在 profile 的 package.json 的 dsh.profile.bundles 中追加 @deepseek-ai/dsh-headless），否则 dsh --profile <P> "任务" 不会启动可回答的 agent。

🚀 使用

dsh --profile <P> "用 skill 工具加载 pdf 技能，把这份文档转成 PDF"

模型会在需要时自行从 catalog 加载对应技能；也可以指名让模型加载某个技能。全部技能内容见 skills/ 目录（与上游目录结构一致）。

❓ FAQ

• 为什么是 884 个 SKILL.md 而不是上游的 897 个？ 上游实际有 897 个技能文件，其中 13 个属 Anthropic source-available（文档技能），其 LICENSE 禁止再散布，已移除（见上方说明）；另有 1 个 postgres-schema-design/SKILL.MD 副档名大写（逐字节内容相同），为在大小写敏感文件系统上可解析，改名 SKILL.md。
• 技能为什么是英文？ 上游即英文；「100% 原样复制」规则禁止翻译技能内容，本 README 才是中文门面。
• 重名技能怎么处理？ 21 个技能名跨类别重复（45 个文件，14 组内容完全相同、7 组不同），dsh catalog 以名称为键只展示一份（类别字母序最前者），所有文件都在 repo 里；被遮蔽的内容可由文件路径取用。
• 「逐字保留」怎么自验？ node scripts/verify-fidelity.mjs，或对照 THIRD_PARTY_NOTICES.md 里的 diff 命令。

🔧 移植改动（完整清单见 THIRD_PARTY_NOTICES.md）

1. SKILL.MD → SKILL.md（大小写敏感文件系统可解析，内容不变）
2. torch_geometric 注册名用其 frontmatter 自声明的 torch-geometric（registry 要求 kebab-case）
3. 递归收集嵌套技能（如 game-development/2d-games，上游本身如此组织）
4. 新增 dsh 载入外壳（lib/ + package.json + cordis.patch.yml），采用与 dsh-lens 相同的 provider 写法

📜 License

本 repo 的 MIT 只涵盖本移植自己的代码：载入外壳（lib/）、打包（package.json、cordis.patch.yml）、测试（test/、scripts/）。双版权：上游 Copyright (c) 2025 Daniel (San) Ávila (claude-code-templates) + Copyright (c) 2026 GongYuanCaiJi (dsh port)。package.json 的 license: "MIT" 指的就是这一部分。

skills/ 与 .claude-plugin/ 底下的上游内容不是统一 MIT，而是混合授权，以各目录自己的 LICENSE 为准：56 份目录级 LICENSE = 44 份 Apache-2.0 + 12 份 MIT（版权归各作者）。原先随上游带入的 13 份 Anthropic source-available（© 2025 Anthropic, PBC. All rights reserved）因 LICENSE 明文禁止再散布，已从本 repo 移除；需要请直接去上游（anthropics/skills 或 davila7/claude-code-templates）取用。逐目录明细见 THIRD_PARTY_NOTICES.md；Anthropic 的完整授权说明见 ANTHROPIC_ATTRIBUTION.md（上游原样保留，也见上游 LICENSE）。

如果你喜欢这套技能，请也给上游 davila7/claude-code-templates 一个 star ⭐

English

All 884 Claude Code skill templates, ported wholesale to DeepSeek Harness (dsh) — nothing filtered, everything shipped.

A dsh port of davila7/claude-code-templates (MIT, 30k+ stars). The 884 SKILL.md files shipped from upstream (897 total upstream, minus 13 Anthropic source-available skills removed for licensing — see below) and all 4160 companion files (references/, examples/, assets/, …) are copied byte-for-byte; skill content is untranslated and unmodified, wrapped in a minimal dsh loading shell.

13 Anthropic source-available directories removed. Upstream's skills/ANTHROPIC_ATTRIBUTION.md marks these document skills as source-available (reference only, not redistributable), and their LICENSE.txt explicitly forbids Reproduce / Distribute / sublicense / transfer to any third party; public upstream hosting is not a license for us to redistribute — each distributor is independently responsible. So document-processing/{docx, docx-official, pdf, pdf-anthropic, pdf-official, pptx, pptx-official, xlsx, xlsx-official} and scientific/document-skills/{docx, pdf, pptx, xlsx} have been removed from this repo; all Apache-2.0 / MIT content is retained. Get these skills directly from upstream: anthropics/skills or davila7/claude-code-templates.

✨ Features

• All 884 skills shipped: the dsh skill catalog shows 860 unique skills (21 names repeat across categories, 45 files total; 14 pairs byte-identical, 7 differ; the catalog keys by name and shows one winner — all files remain in the repo)
• 29 categories: ai-research, development, business-marketing, creative-design, scientific, security, web-development, workflow-automation, and more — including Anthropic's Apache-2.0 official skills (attribution preserved upstream, kept verbatim)
• Verifiable fidelity: THIRD_PARTY_NOTICES.sha256 pins SHA-256 of all 5044 files; npm prepare / npm prepack fail on drift
• Includes the official Claude plugin skill .claude-plugin/skills/owasp-security (OWASP six-standard security reference with examples/)
• Model-invocable through dsh's standard skill tool, zero extra configuration

📸 Demo

Skill catalog in a headless session (real output, excerpt):

- `2d-games`: 2D game development principles. Sprites, tilemaps, physics, camera.
- `ab-test-setup`: When the user wants to plan, design, or implement an A/B test or experiment...
- `git-commit-helper`: Generate descriptive commit messages by analyzing git diffs...

The model loads skills through the skill tool and follows them:

$ dsh --profile <P> "Load the skill named git-commit-helper and quote its first heading."
# Git Commit Helper

📦 Install

The repo is private pending human review; use a local path, or GitHub once public:

# Local path (not yet published to npm or open-sourced)
dsh plugin --profile <P> add /path/to/dsh-claude-code-templates

# GitHub (after open-sourcing; allow build scripts when prompted)
dsh plugin --profile <P> add github:GongYuanCaiJi/dsh-claude-code-templates

Note: a fresh profile needs the headless mode bundle — dsh plugin --profile <P> add @deepseek-ai/dsh-headless and append @deepseek-ai/dsh-headless to dsh.profile.bundles in the profile's package.json — otherwise dsh --profile <P> "task" will not boot an answering agent.

🚀 Usage

dsh --profile <P> "Use the skill tool to load the pdf skill and convert this document to PDF"

The model loads the matching skill from the catalog on demand. All skill content lives under skills/, mirroring the upstream directory structure.

❓ FAQ

• 884 SKILL.md, not upstream's 897? Upstream ships 897 skill files; 13 are Anthropic source-available document skills whose LICENSE forbids redistribution, so they are removed (see above). One more (postgres-schema-design/SKILL.MD) has an uppercase extension (identical bytes) — renamed to SKILL.md so case-sensitive filesystems can resolve it.
• Why are the skills in English? The upstream is English; the "100% verbatim copy" rule forbids translating skill content. This README is the Chinese-facing surface.
• Duplicate names? 21 names repeat across categories (45 files; 14 pairs identical, 7 differ). The dsh catalog keys by name and shows one winner (first alphabetically). All files remain in the repo; shadowed content is reachable by file path.
• How do I verify fidelity? node scripts/verify-fidelity.mjs, or the diff commands in THIRD_PARTY_NOTICES.md.

🔧 Port changes (full list in THIRD_PARTY_NOTICES.md)

1. SKILL.MD → SKILL.md (case-sensitive FS resolvability; bytes unchanged)
2. torch_geometric registered under its frontmatter-declared torch-geometric (registry requires kebab-case)
3. Recursive collection of nested skills (e.g. game-development/2d-games, as upstream organizes them)
4. Added the dsh loading shell (lib/ + package.json + cordis.patch.yml), using the same provider pattern as dsh-lens

📜 License

This repo's MIT covers only this port's own code: the loading shell (lib/), packaging (package.json, cordis.patch.yml), tests (test/, scripts/). Dual copyright: upstream Copyright (c) 2025 Daniel (San) Ávila (claude-code-templates) + Copyright (c) 2026 GongYuanCaiJi (dsh port). The license: "MIT" field in package.json refers to that part only.

The upstream content under skills/ and .claude-plugin/ is not uniformly MIT — it is mixed-licensed, and each directory's own LICENSE governs: 56 directory-level LICENSE files = 44 Apache-2.0 + 12 MIT (copyright to their respective authors). The 13 Anthropic source-available directories (© 2025 Anthropic, PBC. All rights reserved) that originally came with the upstream have been removed from this repo because their LICENSE forbids redistribution; fetch them from upstream (anthropics/skills or davila7/claude-code-templates) directly. Per-directory breakdown in THIRD_PARTY_NOTICES.md; Anthropic's full attribution statement is preserved verbatim in ANTHROPIC_ATTRIBUTION.md (see also the upstream LICENSE).

If you like these skills, please also star upstream davila7/claude-code-templates ⭐