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
dsh-claude-code-templates · DSH Hub
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
"Load the skill named git-commit-helper and quote its first heading."
# Git Commit Helper
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
"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
add /path/to/dsh-claude-code-templates
# GitHub (after open-sourcing; allow build scripts when prompted)
dsh plugin --profile
add github:GongYuanCaiJi/dsh-claude-code-templates
Note: a fresh profile needs the headless mode bundle — dsh plugin --profile add @deepseek-ai/dsh-headless
and append @deepseek-ai/dsh-headless to dsh.profile.bundles in the profile's package.json —
otherwise dsh --profile "task" will not boot an answering agent.
🚀 Usage
dsh --profile
"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)
torch_geometric registered under its frontmatter-declared torch-geometric (registry requires kebab-case)
Recursive collection of nested skills (e.g. game-development/2d-games, as upstream organizes them)
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.
