literaf
dsh-research
AI4Scholar research pack for DeepSeek Harness (dsh): embedded research skills, literature-first workflow guidance and workspace conventions. Pairs with dsh-ai4scholar.
- Stars
- 0
- Language
- TypeScript
- Created
- Aug 16, 2026
- Updated
- Aug 16, 2026
Introduction
dsh-research
AI4Scholar Research Pack for DeepSeek Harness
English | 中文
The orchestration layer that turns DeepSeek Harness into a research workbench. It searches nothing itself: it ships the research skills, the literature-first workflow guidance, and the workspace conventions that make capability plugins (dsh-ai4scholar) and the agent behave like one research assistant.
What it installs
Four research skills (authored by AI4Scholar, embedded in the package — nothing to download):
| Skill | When the agent loads it |
|---|---|
ai4scholar-paper-review | Pre-submission review from five reviewer perspectives, with a fix-priority list |
ai4scholar-introduction-writing | Introductions: background → gap → contribution → signposting → polish |
ai4scholar-academic-formatting | Reference formatting, three-line tables, abstract structure, re-submission reformatting |
ai4scholar-reference-audit | Reference audit: cross-citation, format consistency, DOI checks, duplicates, self-citation |
One workflow section in the system prompt: search before asserting, which task loads which skill, and a workspace layout — papers in papers/, citations in references.bib, reading notes in notes/ — so the next session can pick up where the last one stopped.
What the agent does with it
A request like "find the strongest recent evidence for X, then draft an introduction around it" runs as one chain instead of four disconnected answers:
- The guidance forbids answering from memory, so the agent calls
search_papersfirst and works from what came back. - It reads the full text of the papers that matter with the
read_*tools, and files them underpapers/with one note each innotes/. ai4scholar-introduction-writingloads before any prose is written, so the draft follows background → gap → contribution → signposting rather than improvised paragraphs.- Every work it cites lands in
references.bib, andai4scholar-reference-auditcan check that list against the text before submission.
The next session starts from those three files, so the work accumulates instead of restarting.
How it adapts
The guidance is rebuilt at every prompt assembly, so it states what is actually available right now:
| Composition | What the pack contributes |
|---|---|
dsh-research + dsh-ai4scholar | Skills, plus a workflow that routes the model to the literature tools by name |
dsh-research alone | Skills, plus a workflow that tells the user how to install the literature tools before relying on any citation |
| A profile without a skill registry | The workflow only. It stops advertising the skills rather than sending the model after a loader that cannot resolve them |
Install
dsh plugin --profile web add dsh-ai4scholar dsh-research
dsh web
One command, two independent layers: dsh-ai4scholar brings 38 literature tools (Semantic Scholar / PubMed / Google Scholar / arXiv / bioRxiv / DOI full text / auto-cite / figures), this pack brings the skills and the workflow. The pack works alone too — the skills still load, there is just nothing to search with.
Configuration
The bundle inserts one row (id: research-pack). Override it from your profile's cordis.patch.yml (a patch replaces the whole config, so restate every key you keep):
- id: research-pack
config:
skills: true # register the bundled skills
skillNames: [] # register only these; empty means all
promptGuidance: true # register the workflow section
promptOrder: 140
workspaceConventions: true # state the workspace layout in the guidance
papersDir: papers/
bibliographyFile: references.bib
notesDir: notes/
language: zh # zh | en — guidance and skill catalog copy
Language
language selects the language of the workflow section and of each skill's catalog entry. The skill bodies stay Chinese in both settings — they are AI4Scholar's originals, kept verbatim so they stay in sync with their published source. Under language: en the guidance says so explicitly and tells the model to carry the procedure out but report back in the language the user is writing in.
Notes
- Skill bodies are the ones published on ai4scholar.net, embedded through
ctx.skills.register— no directories to configure, and written in Chinese (see Language). - This pack is an orchestration layer, not a container: member plugins install beside it and keep their own settings pages, cards, and release cadence.
- Skill content is MIT licensed; use and adapt freely.
License
MIT