Research Paper Kb

Persistent cross-session knowledge base for research papers. Ingest arXiv/DOI → extract method, gap, threat level → append to PAPERS.md. Never lose paper con...

Persistent research paper knowledge base for AI agents.

Ingest any paper (arXiv URL, DOI, or title) and extract structured intelligence into a permanent PAPERS.md knowledge base that survives across sessions. Never lose context on a paper again.

---

Trigger Conditions

Use this skill when the user:

• Pastes an arXiv URL (e.g. https://arxiv.org/abs/2310.12345)
• Pastes a DOI (e.g. 10.1038/s41586-024-07156-8)
• Says "add this paper to my KB" / "track this paper" / "save this paper"
• Says "what do we know about [paper title]"
• Says "update my paper KB" / "scan my PAPERS.md"
• Says "show me the papers I'm tracking" / "what papers have I saved"

---

What This Skill Does

1. Fetches the paper abstract, metadata, and key sections
2. Extracts structured intelligence (method, gap, threat level, overlap)
3. Generates a clean BibTeX entry
4. Appends a structured entry to PAPERS.md in the workspace
5. Updates MEMORY.md with a pointer so future sessions know the KB exists
6. Works across sessions — the knowledge base is a file, not context

---

Step-by-Step Instructions

Step 1: Identify the Paper

Accept any of:

• arXiv URL: https://arxiv.org/abs/XXXX.XXXXX
• arXiv ID: 2310.12345 or 2310.12345v2
• DOI: 10.XXXX/...
• Title string: look up via Semantic Scholar API

Normalize to arXiv ID or DOI before proceeding.

Step 2: Fetch Paper Metadata

For arXiv papers — fetch the abstract page:

https://arxiv.org/abs/<arxiv_id>

Extract: title, authors, date, abstract, subject categories.

Also fetch the Semantic Scholar API for structured metadata:

https://api.semanticscholar.org/graph/v1/paper/arXiv:<arxiv_id>?fields=title,authors,year,abstract,tldr,citationCount,influentialCitationCount,fieldsOfStudy

For DOI papers — use Semantic Scholar:

https://api.semanticscholar.org/graph/v1/paper/<DOI>?fields=title,authors,year,abstract,tldr,citationCount,influentialCitationCount,externalIds

For title lookup:

https://api.semanticscholar.org/graph/v1/paper/search?query=<url_encoded_title>&fields=title,authors,year,abstract,externalIds&limit=1

Step 3: Extract Structured Intelligence

From the abstract and any available full text, extract:

Field	What to Extract
Method	Core technical approach or contribution (1-2 sentences)
Gap they claim	What problem/limitation they say they're solving
Key results	Main quantitative or qualitative outcome
Overlap with user's work	Ask the user if context is unclear; or infer from prior PAPERS.md entries and MEMORY.md
Threat level	1-5 scale: how much does this threaten the user's research? (1=unrelated, 5=directly competing)
Citation count	From Semantic Scholar
Related papers	Up to 3 highly-cited related papers from the same fetch

Threat level guide:

• 1 — Unrelated field, no overlap
• 2 — Adjacent method, different application
• 3 — Similar approach, different dataset/domain
• 4 — Direct competition, overlapping claims
• 5 — Near-identical work, same target problem

Step 4: Generate BibTeX

Generate a clean BibTeX entry. Format:

For arXiv:

@article{<AuthorYEARkeyword>,
  title     = {<Full Title>},
  author    = {<Author1> and <Author2> and ...},
  journal   = {arXiv preprint arXiv:<id>},
  year      = {<year>},
  url       = {https://arxiv.org/abs/<id>},
  note      = {arXiv:<id>}
}

For published paper:

@article{<AuthorYEARkeyword>,
  title     = {<Full Title>},
  author    = {<Author1> and <Author2> and ...},
  journal   = {<venue>},
  year      = {<year>},
  doi       = {<doi>},
  url       = {https://doi.org/<doi>}
}

BibTeX key convention: FirstAuthorLastNameYearFirstContentWord (e.g., Smith2024diffusion)

Step 5: Write to PAPERS.md

Check if PAPERS.md exists in the workspace root. If not, create it with the header:

# PAPERS.md — Research Paper Knowledge Base
> Auto-maintained by the `research-paper-kb` skill. Add papers with: "add this paper to my KB"
> Last updated: <date>

---

Append (never overwrite) the following entry template:

## [<Short Title>](<arxiv_or_doi_url>)
**Added:** <YYYY-MM-DD>  
**Authors:** <Author1>, <Author2>, ...  
**Venue:** <arXiv / Conference / Journal>  
**Citations:** <N> (Semantic Scholar)  
**Threat Level:** <1-5> — <one-line reason>

### Method
<1-2 sentence description of the core technical contribution>

### Gap They Claim
<What problem/limitation they say they're solving>

### Key Results
<Main outcomes, benchmarks, or claims>

### Overlap With My Work
<How this relates to the user's research — ask if unclear>

### Notes
<Any additional context the user provides, or leave blank>

### BibTeX
```bibtex
<bibtex entry>

---

### Step 6: Update MEMORY.md

After writing to PAPERS.md, append or update the PAPERS.md pointer in `MEMORY.md`:

Find or create a section `## Research Paper KB`:
```markdown
## Research Paper KB
- PAPERS.md exists in workspace root — <N> papers tracked as of <date>
- Latest addition: <Short title> (<threat level>/5)
- Run `research-paper-kb` to add more papers

If the section already exists, update the count and latest addition line.

Step 7: Confirm to User

Reply with a summary:

✅ Added to PAPERS.md

**[Paper Title]** (<year>)
- Authors: ...
- Threat level: X/5 — <reason>
- BibTeX key: `AuthorYearWord`

PAPERS.md now has N papers. Run `show me my papers` to review.

---

Query Mode: "Show Me My Papers"

When the user asks to review their paper KB:

1. Read PAPERS.md
2. Summarize: total count, highest threat-level papers, recently added
3. Optionally filter by threat level, topic, or year
4. Offer to export BibTeX for all papers: collect all @article{...} blocks and present as a code block

---

Query Mode: "What Do We Know About X?"

When the user asks about a specific paper or topic:

1. Search PAPERS.md for matching title/author/keywords
2. Return the structured entry
3. If not found, offer to add it: "This paper isn't in your KB yet. Want me to add it?"

---

Edge Cases

Situation	Handling
arXiv paper not found	Try Semantic Scholar title search; if still not found, ask user to confirm title
DOI behind paywall	Fetch abstract from DOI.org metadata (https://doi.org/<doi> with Accept: application/json); note "full text unavailable"
Paper already in PAPERS.md	Detect by title/arXiv ID match; offer to update notes or threat level instead
User doesn't know their research area	Ask: "What's your research focus? I'll use this to assess overlap." Store in MEMORY.md
Semantic Scholar rate limit	Fall back to arXiv API: http://export.arxiv.org/api/query?id_list=<id>

---

Integration With Other Skills

This skill works best alongside:

• citation-management — for full BibTeX workflow and PubMed/Google Scholar search
• biorxiv-database — for biology/life-science preprints (use to find papers to add)
• cs-research-methodology — for identifying gaps and research proposals from your KB
• proactive-research (ClaWHub) — can feed new papers into this KB automatically

---

Files Modified

File	Action
PAPERS.md	Append new entry (create if missing)
MEMORY.md	Update ## Research Paper KB section

Never modifies: SOUL.md, USER.md, AGENTS.md, TOOLS.md, or any project files.

---

Example Interaction

User: "Add this to my KB: https://arxiv.org/abs/2310.06825"

Agent:

1. Fetches arXiv 2310.06825 → "Mistral 7B" by Jiang et al.
2. Fetches Semantic Scholar metadata (12k citations)
3. Extracts: method = grouped-query attention + sliding window; gap = efficient 7B model
4. Assesses threat level vs user's work (reads MEMORY.md for context)
5. Generates BibTeX key Jiang2023mistral
6. Appends structured entry to PAPERS.md
7. Updates MEMORY.md
8. Replies: "✅ Added Mistral 7B (threat: 2/5 — efficient inference, different from your focus on X)"

---

Metadata

name: research-paper-kb
version: 1.0.0
author: <your-github-handle>
category: Academic & Research
tags: [research, papers, arxiv, bibtex, knowledge-base, literature, academic, persistent-memory]
summary: Persistent cross-session knowledge base for research papers. Ingest arXiv/DOI → extract method/gap/threat level → append to PAPERS.md. Never lose paper context again.
requires: []