Introduction

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin,” aiming to extend agent capabilities through modular expansion. To address common problems in academic writing by agents—such as the lack of structured assistance, non-standard citation formatting, and missing language style validation—the dsh-plugin-academic-writing plugin was created. It does not generate final paragraphs; instead, it provides the model with paper skeletons, formatting standards, and deterministic validation logic, helping developers build more standardized agent workflows.

What This Is

This is an academic writing toolkit for DeepSeek Harness agents. It is maintained by user 863683348 and is released under the MIT license. Its core value is packaging structure planning, citation formatting, LaTeX code generation, and language style validation into reusable tools, solving the problem of agents lacking standard templates and validation capabilities when writing papers.

Core Features

The plugin provides ten core tools, covering the full process of paper writing:

  1. Outline skeleton (outline): Generates a paper outline, supporting types such as research, review, proposal, and essay.
  2. Title variants (title): Generates titles based on style templates (descriptive, interrogative, declarative, colon-style).
  3. Abstract skeleton (abstract): Generates a structured abstract framework (background, methods, results, conclusion).
  4. Citation formatting (citation): Formats reference entries according to GB/T 7714, APA 7, or MLA 9 standards.
  5. Language style validation (check): Checks text for passive voice, first-person usage, hedging words, and sentence length.
  6. Pre-submission checklist (checklist): Provides a preliminary checklist covering ethics, conflicts of interest, data, funding, and formatting.
  7. Reference list (references): Generates a reference list from entries, with support for deduplication and sorting.
  8. LaTeX snippets (latex): Provides booktabs tables, formulas (inline / display / aligned), code listings, and the thebibliography environment.
  9. Manuscript structure check (structure): Checks heading levels, paragraph length, and outline coverage.
  10. Glossary (glossary): Generates a glossary of terms and abbreviations, including initial-appearance definitions.

Installation and Enablement

Installing the plugin requires the DSH CLI tool:

dsh plugin --profile <profile> add dsh-plugin-academic-writing

After installation, DSH must be restarted to register the academic_writing tools globally.

Typical Usage

In DSH’s Composition Row, behavior can be adjusted via the config parameter. The default values are:

{
  "config": {
    "personaSection": true,
    "sectionOrder": 6
  }
}
  • personaSection: Defaults to true, used to register the academic prompt guidance section.
  • sectionOrder: Defaults to 6, controlling the order of prompt sections (persona is 0, in ascending order).

When invoked, the agent can use the aforementioned tools, such as outline, title, and citation, to collaborate.

Design Principles and Notes

The core logic of the plugin is located in lib/academic.js, a pure-logic file that does not depend on DSH or Cordis imports and is covered by independent unit tests. lib/index.js is only a thin wrapper for the Cordis plugin.

  • Compatibility: Tool schemas have been validated with the @deepseek-ai/dsh-tools value-schema DSL. Early versions were incompatible with the value-schema DSL when handling the required property in JSON schemas, which could cause DSH startup failure (error: unsupported JSON schema: schema.required is not supported by the value schema DSL). The current version has fixed this issue. If an affected version prevents DSH from starting, remove the plugin or upgrade the version; no data will be lost.
  • Permissions and side effects: At runtime, the plugin has no filesystem access permissions, and all operations are deterministic and side-effect free.

Ecosystem and Closing

DeepSeek Harness emphasizes a plugin-based ecosystem. This plugin is listed in the community directory SkillHub and is not affiliated with the official DeepSeek or Huanfang.