| name | arxivsub-skill |
| description | Use whenever the user wants to search for academic papers, find recent research, look up conference publications, or explore literature on any AI / ML / CV topic. Routes through the `arxivsub-search` MCP server (arXiv + CVPR / ICCV / ICLR / ICML / NeurIPS / AAAI / MICCAI). Triggers on: "find papers on X", "what are the latest papers about Y", "search arXiv for Z", "any recent work on W", '论文检索', '最新研究'. |
| version | 0.2.0 |
arxivsub-skill
Search academic papers via the arXIVSub API through MCP.
Language Rule
Always respond in the same language the user is using.
Step 0: Authentication
The arxivsub-search MCP server reads the API key automatically from the environment or a .env file at the workspace root. Never ask the user for it unless the MCP call fails with a missing-key or auth error, and never pass it as a chat parameter.
If the MCP tool reports missing_api_key or an auth failure involving ARXIVSUB_SKILL_KEY, tell the user (in their language) to set it up via one of:
- Export as a shell environment variable (add to
~/.zshrc or ~/.bashrc for persistence):
export ARXIVSUB_SKILL_KEY=your_key_here
- Add to a
.env file in the working directory:
ARXIVSUB_SKILL_KEY=your_key_here
The user's API key is found on the Skills page of the arXivSub website.
Step 1: Show search parameters and execute
Before calling the API, briefly show the interpreted parameters in one line (in the user's language), then proceed without waiting:
Searching: query="...", locations=[...], time=..., limit=...
Pause only if the search intent is genuinely ambiguous (e.g. a term that could mean multiple very different topics).
Step 2: Call MCP search
Call arxivsub-search.search_papers with the interpreted parameters. Omit arxiv_days / conference_years if not applicable.
Expected MCP arguments:
{
"query": "<search terms>",
"locations": ["arxiv", "CVPR", "NeurIPS"],
"limit": 10,
"arxiv_days": 7,
"conference_years": [2024, 2025],
"language": "en"
}
The MCP tool returns full paper details and quota_remaining in structured content. Do not create temp JSON files.
Step 3: Filter and rank
From the returned list, select the top 5–10 using:
- Relevance first — how directly the paper addresses the query
- Recency as tiebreaker — among equally relevant papers, prefer the most recent
Step 4: Fetch full details and respond
Compose the response from the returned details. Never mention files, scripts, temp files, or internal mechanics.
Output structure (translate headers to the user's language)
[Research Findings] — Synthesize insights. Answer the user's question directly.
[Recommended Papers] — For each paper, write a substantive description (not just what_about). Cover: what problem it solves, the key method or contribution, and notable results or significance. Typically 3–5 sentences.
**[Title]**
📍 [conference / arXiv] · [year if available]
👥 [first_author] ([first_aff]) · [last_author] ([last_aff])
📄 [synthesized description based on full paper details]
🔗 [pdf_url]
At the bottom, show quota in the user's language as a footnote:
English: Daily quota remaining: N searches / Chinese: 当日剩余搜索额度：N 次
Key rules
locations is case-sensitive: arxiv, CVPR, ICCV, ICLR, ICML, NeurIPS, AAAI, MICCAI- Show parameters as a one-liner before calling the MCP; only ask for confirmation if intent is ambiguous
- If the MCP returns an error, classify and handle:
- Retryable (network timeout, transient server error): inform the user; offer to retry
- Needs user intervention:
- Quota exhausted → tell the user the daily quota is used up; do not retry
- Auth failure / missing key → go to Step 0
- Empty results → suggest broadening the query (fewer locations, wider date range, looser terms)
- JSON parse failure / malformed response → report as an unexpected error; ask the user to try again later
- NEVER output raw JSON or expose internal mechanics
- NEVER call local skill scripts; all network access MUST go through MCP
