Preface

When using Large Language Models (LLMs) for scientific research, a common gap is that after a round of chat, the answer remains only in the chat history. Literature reviews cannot be archived, citations cannot be verified, and switching models requires re-explaining the context. For developers in the DeepSeek Harness (DSH) ecosystem, DSH’s philosophy is “everything is a plugin”; what is often missing is not model capability, but a reusable and verifiable research workflow written into a skill.

The shidi-skill introduced below is a skill package designed around this problem: fully executing tasks like literature research, experimental design, plotting, and paper deep reading, and delivering real files to disk.

What is it

shidi-skill is a skill package for the AI-for-Science research workflow in DeepSeek Harness, maintained by IcyCreamDAS, under the MIT license. Its positioning can be summarized in one sentence: covering literature research / experimental design / plotting / paper deep reading, delivering real files, mandating the output of cross-validation task sheets, and having zero external dependencies.

Two design principles are worth mentioning separately:

  1. The deliverables are files, not chat text. The final report is written to ~/shidi-output/, and the cross-validation task sheet to ~/shidi-verify/, allowing for direct archiving, citation, or forwarding.
  2. Every delivery must be accompanied by a cross-validation task sheet. You can hand the task sheet to another LLM for independent verification, respond to verification comments item by item, and iterate in a loop.

In terms of compatibility, it follows the SKILL.md specification. The README indicates it can be used with agents such as Claude Code / Codex / OpenClaw / OpenCode, and is not tied to a single model or platform. The version number is recorded in two places: the README badge shows 0.3, and in package.json it is 0.3.1.

Core Capabilities

According to the capability table in the README, it covers six types of tasks, starting with the main workflow.

Literature Research: Confirm three parameters (quantity, domain, special requirements) at once and start; diverge into 3×3=9 angles, perform targeted searches in top journals and follow a fallback chain; perform a six-dimensional rating on results, then read them by level; subsequently download PDFs in batch, and finally output a file report.

Experimental Design: First freely diverge on solutions, then support them with literature, and output principles, steps, precautions, and a list of loopholes/vulnerabilities—the list is explicitly listed without concealment.

Scientific Plotting: Read from CSV/TXT/Excel, clean first then calculate (numpy/scipy/sympy), render publication-grade figures, and attach Origin replication instructions.

Paper Deep Reading: source-map six-step deep reading, producing 16 paper cards and a glossary.

Cross-Validation Gate: The author calls this the soul of the system. Every delivery must produce a verification task sheet; after handing it to another LLM for independent verification, it iterates in a loop.

Experiment Logging: Standardize Experiment ID and Batch ID, and archive using YAML frontmatter.

The most recent batch of entries in the README changelog (2026-08) added four capabilities:

  • PDF Batch Download: OpenAlex/Unpaywall automatically locates OA direct links; if it fails, it generates a categorized report instead of silently skipping.
  • Multi-format Report: Markdown → HTML → PDF (Headless Edge printing, zero dependencies).
  • PDF Extraction Layer: Converts text-based PDFs to Markdown to save context; scanned documents are automatically tagged as needs_vision to take the vision path.
  • Angle-specific Intermediate Reports: When researching from multiple angles in parallel, a separate file is saved for each angle, while the summary is a read-only file.

Safeguard Mechanisms

The biggest risk for research tasks is fabricating citations. shidi-skill has built-in several hard rules for this:

  1. A cap on each dimension of the six-dimensional rating, with the total score recalculated.
  2. A veto if topic matching is below 10 points.
  3. Three-level deduplication using DOI / arXiv ID / URL.
  4. No fabrication: any unverified citation, DOI, or data is labeled as “unverified”.
  5. Transparent reading depth: Full text ✅ / Abstract only ⚠ / Unavailable ❌
  6. Respond to cross-validation comments item by item (with reasons for acceptance or rejection), without blindly following.

Installation and Usage

The installation method given in the README is to clone and copy to the skills directory, which loads the skill into the local agent skills path:

git clone https://github.com/IcyCreamDAS/shidi-skill.git
cp -r shidi-skill/skills/shidi ~/.claude/skills/

No official DSH installation commands such as dsh plugin add were found in the documentation; the two lines above are all the installation steps recorded in the README.

The plugin’s package.json declares DSH-related metadata; the key snippet is as follows:

{
  "name": "shidi-dsh-plugin",
  "version": "0.3.1",
  "engines": { "node": "^22.19.0 || >=24" },
  "peerDependencies": {
    "@deepseek-ai/dsh-skill-filesystem": ">=0.1.0-rc.6 <0.2.0"
  }
}

Additionally, the dsh.bundle.patch field in package.json points to ./cordis.patch.yml.

Dependencies are divided into two layers: the core is pure SKILL.md and Markdown, with zero external dependencies, requiring no package installation; optional enhancements are in two places: the plotting branch uses numpy/scipy/matplotlib/pandas/sympy (pyvista for 3D plots), and the PDF extraction layer uses pdf-inspector. Literature search follows a fallback chain: falling back to open academic APIs (OpenAlex / Crossref / arXiv) after WebSearch, all without requiring an API key.

Typical Usage

After installation, you can simply issue tasks using natural language. The examples given in the README are as follows:

师弟，帮我查一下 XX 主题的文献，要 20 篇
师弟，设计一个 XX 实验方案
师弟，把这个数据画成图
师弟，帮我把这篇论文精读一下

You can just shout “师弟” (Junior Brother); it will confirm the task type with you before starting. The persona only takes effect when the trigger word “师弟” appears; the user is addressed as “师兄” (Senior Brother) by default.

Taking literature research as an example, after the above steps, the complete workflow is: confirm three parameters → search by 9 angles (saving an independent intermediate report for each angle) → six-dimensional rating and three-level deduplication → graded reading and batch PDF download → generate report in ~/shidi-output/ and verification task sheet in ~/shidi-verify/. After obtaining the task sheet, you can hand it to another LLM for independent cross-validation; it will respond to verification comments item by item and then iterate.

Applicable Scenarios and Notes

Suitable audience: STEM graduate students, people doing AI for Science, and agent users who need archivable and cross-verifyable file outputs. If your daily work revolves around literature search, design, plotting, and reading papers, this skill covers the entire workflow.

几点注意：

  1. The plugin runs with the permissions of the current DSH process and can read/write files on your machine; check the source code and license (MIT, LICENSE included in the repo) before installing.
  2. The version number in the README badge (0.3) and package.json (0.3.1) are inconsistent; rely on the actual code in the repository.
  3. There are two naming schemes: the directory identifier is icycreamdas/shidi-skill, and the package name in package.json is shidi-dsh-plugin; pay attention to distinguish them when referencing.
  4. Zero dependencies only applies to the main workflow; optional enhancements for plotting and PDF extraction require corresponding Python packages, install as needed.

Summary

The value of shidi-skill is not in a single point function, but in writing the entire chain of “ask parameters — search — rate — read — output file — cross-verify” into a skill, and retrieving results from chat history using disk files and verification task sheets. For developers who want to build a research workflow on DSH, it is also a reference implementation that can be read directly and modified as needed.

  • GitHub Repository: https://github.com/IcyCreamDAS/shidi-skill
  • Community Directory Page: https://www.skillhub.cn/plugins/IcyCreamDAS/shidi-skill

Note: skillhub.cn is an independently maintained community directory site and has no official affiliation with DeepSeek or High-Flyer.