Introduction¶
The philosophy of DSH is “everything is a plugin”. For agent developers, a common pain point is not that a single prompt is insufficient, but that tasks such as paper reading, diagramming, document conversion, and security auditing are difficult to reuse across projects. Azzygoatcoder/agent-useful-skills organizes these tasks into skill + script modules and is compatible with Claude Code / DeepSeek Harness. This introduces it within the DSH plugin ecosystem and does not present the community directory as an official app store.
What is this¶
Azzygoatcoder/agent-useful-skills is a modular collection of AI research/engineering skills maintained by Azzygoatcoder, licensed under MIT. Its goal is to solidify repetitive processes like “reading papers, drawing diagrams, writing documentation, and security auditing” into registerable, verifiable, and deployable skill modules. The documentation states that the design inspiration comes from ARIS, but its code is not used directly; attribution for third-party content can be found in THIRD-PARTY-NOTICES.md.
Core Features¶
The following introduces several verified categories of capabilities.
Security Auditing¶
Supports scenario routing, parallel exploration, deep verification, reporting, incremental re-auditing, and state tracking. In actual use, you can trigger the code-security-audit skill or use /audit directly.
Paper Workflow¶
Covers paper reading, paper writing, LaTeX template management, and compilation page count checking.
Office Processing¶
Supports markdown to docx/pptx conversion, Excel processing, and image extraction.
Data and Vector Graphics¶
Supports journal-quality data plots, draw.io replication, and vector graphics consistency checking.
Read-Only Disk Storage Analysis¶
Provides read-only disk storage analysis and interactive HTML reports.
Git Collaboration and Release¶
Covers issues, pull requests (PRs), releases, and reviews.
Cross-Model Validation Loop¶
Provides image rendering verification and cross-model adversarial review, used for secondary checks on generated results.
DSH Plugin Shell and Deployment Scripts¶
Provides DSH plugin shell, skill registration, checking, and deployment scripts. The default registered skill list currently contains 18 skills, and archived skills are not registered by default.
Installation and Activation¶
DSH Plugin Installation¶
The following command is used to install this plugin via DSH:
dsh plugin --profile web add github:Azzygoatcoder/agent-useful-skills
Plugin scripts will run with the permissions of the current dsh process. It is recommended to check the source code, dependencies, and MIT license before installing.
DSH Skill Directory Rules¶
DSH only recognizes single-layer skill directories:
<skill_root>/<skill_name>/SKILL.md
Skills within the plugin need to have junctions created individually; do not link the entire skills/ directory.
Skill Linking Examples¶
Claude Code users can link a single skill to ~/.claude/skills/:
ln -s "$(pwd)/superpowers/skills/paper-reading" ~/.claude/skills/paper-reading
On Windows, DSH can use junctions to point to the user-level skill root:
New-Item -ItemType Junction -Path "$env:USERPROFILE\.dsh\skills\paper-reading" -Target "$pwd\plugins\superpowers\skills\paper-reading"
Under linked deployment, only enable/disable in the DSH skill management interface; do not use delete, as deletion may cascade to the linked target.
Script Invocation¶
Scripts under bin/ can be called directly with Python or executed first:
pip install -e .
After the above steps, you can use:
review file.md
vision img.png "describe this image"
office-tools md2docx a.md b.docx
latex-build list
Typical Usage¶
Image Recognition¶
First configure the environment variables, then execute:
python bin/vision.py <image_path> "describe this image"
The image recognition script depends on LLM_API_URL and API key environment variables. VISION_PROVIDER can switch the image recognition backend, defaulting to siliconflow.
Markdown to Word¶
python bin/office_tools.py md2docx notes.md report.docx --toc
Data Plotting¶
python bin/data_plot.py demo
Security Auditing¶
Trigger the code-security-audit skill, or use /audit directly.
Skill Verification and Deployment¶
Verify that SKILL.md conforms to DSH/AgentSkills rules:
python bin/check_skills.py
Automated deployment/healing of DSH skill links:
pwsh bin/redeploy-skills.ps1
White-box Self-Check¶
node bin/verify-plugin.mjs
Before execution, ensure that node_modules/@deepseek-ai/dsh-skill-filesystem at the repository root is resolvable.
Configuration and Dependencies¶
Scripts prioritize reading environment variables, with a fallback to ~/.claude/settings.json. The repository does not hardcode any vendor endpoints or keys. Taking the default image recognition backend as an example, you can set it up as follows:
export LLM_API_URL="<your API endpoint>"
# Set the corresponding API key environment variables as required by the script
export VISION_PROVIDER="siliconflow"
API keys should only be written in environment variables or local configuration; do not commit them to the repository.
Applicable Scenarios and Notes¶
Suitable for developers who want to solidify repetitive tasks in research and engineering into reusable skills. Notes before use:
- The DSH plugin will run scripts within the repository; it is recommended to check the source code and MIT license before using it.
- The three scripts without file extensions under
subagent-driven-development/scripts/are bash scripts; Windows requires running them in Git Bash / WSL. - Under linked deployment, only enable/disable in the DSH skill management interface; do not delete to avoid cascading effects on the linked target.
- DSH only recognizes single-layer skill directories; junctions for skills within the plugin need to be created individually.
Conclusion¶
The value of agent-useful-skills lies in transforming processes like reading papers, drawing diagrams, writing documentation, and security auditing into registerable, verifiable, and deployable DSH/Claude Code skill modules. GitHub repository: