vista
This skill should be used when the user asks about research directions, open problems, future work, or follow-up research from the academic literature. Trigger phrases include "open problems in X", "what's next for Y", "future work for paper Z", "research directions on T", "salient follow-up research", "what should I work on next in", "find me follow-up research unrelated to my prior work", "broad survey of W", "neglected directions in V". Routes the request to the right Vista MCP tool, reads back the structured paper sections, and synthesizes research directions in the conversation.
- License 许可
- MIT
- 最近更新
- 2026-10-23
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- vista
- description
- This skill should be used when the user asks about research directions, open problems, future work, or follow-up research from the academic literature. Trigger phrases include "open problems in X", "what's next for Y", "future work for paper Z", "research directions on T", "salient follow-up research", "what should I work on next in", "find me follow-up research unrelated to my prior work", "broad survey of W", "neglected directions in V". Routes the request to the right Vista MCP tool, reads back the structured paper sections, and synthesizes research directions in the conversation.
02 资源文件 2
- 参考文档references/output-shape.md4.7 KB
- 参考文档references/tool-routing.md3.0 KB
03 SKILL.md 正文
Vista
Vista surfaces salient research directions by reading the Future Work, Limitations, Discussion, and Conclusion sections of highly-cited papers. The MCP server does the mechanical work (OpenAlex search, PDF fetch, section extraction, caching). This skill tells you when to call which MCP tool and how to synthesize the structured payload it returns.
Core principle
The MCP tools are mechanical. They return raw section text plus metadata. You (Claude in this conversation) read the sections and produce the synthesis: ranking, pattern detection, novelty assessment, cross-paper themes. Do not expect the MCP to "give you the answer"; it gives you the inputs.
When to invoke which tool
Read the user's question carefully and route:
| User intent | Tool | Notes |
|---|---|---|
| "Open problems / future work / research directions in a topic" | topic_followups(topic, max_papers=8, year_min=2020, min_citations=30) |
Most common. Recent, highly-cited, arXiv-available. |
| "What does paper / arXiv:1234.5678 / DOI X say about future work?" | paper_followups(identifier) |
Single-paper deep-dive. Identifier can be DOI, arXiv id, OpenAlex id, or title. |
| "Older / classical / under-explored / neglected directions in X" | find_seminal(topic, year_max=2019, min_citations=1000) |
Legacy track. The "papers everyone cites but nobody followed up on." |
| "Broad / cross-field / serendipitous search; find directions outside my prior work" | broad_followups(query, fields=None, max_papers=12) |
Exploratory; looser thresholds; spans fields. |
| "Search the local catalog of distilled directions" | search_directions(query, ...) |
Use after a build run. Searches what the user has already mined. |
| "List interesting / open / started directions" | list_directions(...) |
Curation/review. |
| "Mark this direction as interesting" | mark_direction(direction_id, status, notes) |
User curation. |
| "Save these directions for paper X" | submit_directions(paper_id, directions) |
After you've synthesized directions, persist them so future runs can search them. Appends by default; curation is never lost. |
| "What papers / directions / fields are in the store?" | status() or run_sql(...) |
Inspection. |
If unsure: prefer topic_followups for topic questions and paper_followups
for paper questions. If the user wants breadth or surprise, use
broad_followups. If the request mentions classical, neglected, or
"never-followed-up", use find_seminal.
Synthesis pattern after a tool call
Each on-demand tool returns {..., "papers": [Paper, ...], "paper_count": N},
where each Paper carries metadata, links, and a sections list of
{type, heading, method, content, char_count} (type is one of
future_work, limitations, discussion, conclusion). The full shapes
are in references/output-shape.md.
For each paper:
- Read the
future_worksection if present, elselimitations, elsediscussion/conclusion(the latter often contain future-work prose without a separate heading). - Pull out concrete, actionable directions. Reject vague aspirations ("more experiments", "scale up").
- For each, capture: a one-sentence direction, a one-line rationale grounded
in the section text, a short verbatim quote (substring of the section), and
tags (
["topic-keyword", "method-keyword"]). - Estimate
feasibilityandnovelty(low/medium/high) based on the paper's stated limits.
Then synthesize across papers:
- Consensus directions: themes that appear across 2 or more papers.
- Frontier directions: novel, mentioned by one paper, well-supported.
- Conspicuous absences: areas the user might expect to be flagged but aren't (use sparingly; it's interpretive).
- Contradictions: papers disagreeing on what to do next.
Present the synthesis with explicit citations to paper title and year, and include the arXiv or DOI URL so the user can drill in.
If the user wants the synthesized directions persisted, call
submit_directions(paper_id, directions) per paper. The default
mode="append" adds new directions and updates ones with the same text in
place (keeping the user's status and notes, merging tags); pass
mode="replace" only to drop the paper's other uncurated directions.
Otherwise leave the catalog alone, since the conversation answer is enough.
paper_followups with a title accepts only an exact (normalized) title. If it
returns "error": "not found" with candidates, show them to the user or
retry with the chosen candidate's paper_id (an OpenAlex id resolves free).
Defaults to nudge
- For ML/AI topics,
topic_followupsdefaults are usually fine. - For statistics, drop
min_citationsto ~20 and setrequire_arxiv=False(many top stats papers are not on arXiv). - For ransomware / security, set
require_arxiv=Falseand lean onmin_citations=15. USENIX/NDSS papers are often OA but not arXiv.
When OpenAlex refuses
If a tool fails with OpenAlexBudgetExhausted (a 429 on a spent daily budget),
do not retry: it cannot succeed until the budget resets. Tell the user that
keyless OpenAlex access shares one small budget per IP, that a free key from
https://openalex.org/settings/api raises it 10x, and that they need to
export OPENALEX_API_KEY=... and restart Claude Code so the MCP server picks
it up. Transient throttling (per-second 429s, 5xx) is retried automatically.
Output style
Default to a concise structured response: synthesis first, then a "directions table" with paper, year, citation count, direction one-liner, and a link. Offer to drill into any direction. Do not paste raw section text back at the user unless they ask for it.
Bulk / pipeline workflows
If the user asks to build a corpus rather than ask an in-conversation
question (e.g., "set up a recurring sweep across all four fields"), use the
bulk tools (discover, extract_pending, render_markdown) or hand them
back to the CLI: vista pipeline --fields ml,ai,stats,ransomware --per-field 25.
Additional resources
references/tool-routing.md: extended decision examples for picking the right MCP tool from ambiguous user phrasings.references/output-shape.md: full JSON schema for each tool's response.
04 同仓库更多 Skills
worldsmith-methodology
Use when working in a project with a worldbuilding documentation ecosystem (docs/ or lore/ directory). Provides the documentation-first methodology: canonical hierarchy, doc roles (timeline, lore, characters, systems), propagation discipline, and the dual canonical and exploratory workflow.
prose-craft
Use when writing fiction, editing fiction prose, reviewing chapters, or drafting scenes. Covers prose craft, dialogue mechanics, scene structure, and common failure modes in AI-assisted fiction. Trigger phrases: "write fiction", "edit chapter", "prose review", "scene draft", "dialogue check", "fiction craft", "edit this scene", "review my prose".
synthesize
Make a research-agent run conclude and write synthesis.md (outcome, key findings, failed approaches, open questions, recommendations). Signals a live researcher to wrap up, or launches one in synthesis mode when none is running. Use when results are good enough or the agent is stalling.
status
Show the current state of a research-agent run from .research/: liveness, cycles and budget, sub-problem and hypothesis statuses, eval trend, current focus, and branches. Read-only.
resume
Resume an interrupted research-agent run from .research/. Re-launches the researcher to reload its structured state and continue from the current focus, optionally on a named branch or with a new budget. Use after a session restart, crash, or explicit pause.
research
Launch an autonomous research agent that iterates toward a goal through proofs, code, simulations, tests, and counterexample search. Fire-and-forget: the agent runs on its own and writes all results to .research/ in the target project.