| name | research-brief-skill |
| description | Lightweight source-backed research and synthesis for normal Reviva agents. Use when the user asks to summarize materials, compare sources, verify claims, collect evidence, create a short research brief, prepare slide/document inputs, or produce a Markdown/HTML brief without running the full deep-researcher pipeline. |
| allowed-tools | file_read, file_write, document_read, kb_search, web_search_bing, mcp:exa |
Research Brief Skill
Use this skill to produce a reliable, compact research brief from local files, knowledge base hits, and optional web search. It is for normal chat agents and task agents that need evidence-backed synthesis without the cost of the full deep-research workflow.
When To Use
Use this skill for:
- summarizing uploaded documents into a cited brief;
- comparing several sources, documents, policies, products, papers, notes, or lesson materials;
- checking whether a claim, number, requirement, exam detail, curriculum point, or public fact has enough support;
- preparing source-backed content for PPT, lesson plans, review notes, office documents, reports, or HTML pages;
- creating a concise Markdown report and optional HTML display page.
Do not use this skill when the user explicitly asks for a long-form deep research report, multi-round investigation, full literature review, or exhaustive market/industry research. Route those tasks to the built-in deep-researcher agent with deep-research.
Core Principle
Prefer trustworthy evidence over volume. A short answer with clear source boundaries is better than a long report that hides uncertainty.
Always distinguish:
- User materials: files or attachments supplied in the current task.
- Knowledge base: cloud KB search results selected by the user.
- Web sources: search results from
web_search_bing, mcp:exa, or another available search tool.
- Model knowledge: general background that is not directly sourced; mark it as such when used.
Source Intake
Handle files by type:
| Type | Tool | Rule |
|---|
.md, .txt, .csv, .json, source snippets | file_read | Read directly, in chunks when large. |
.docx, .pptx, .xlsx | document_read | Use overview first, then text/targeted reads. Never use file_read on Office binaries. |
.pdf | document_read | Use overview first, then page chunks. If unavailable, mark the PDF as not parsed; do not install dependencies. |
| Cloud KB docs | kb_search | Search the KB; do not pretend cloud KB docs have local paths. |
| Web pages/search results | web_search_bing / mcp:exa | Optional. Use only available tools and respect the parent agent's web-search setting. |
Do not auto-install Python packages, ASR models, browsers, scraping tools, or MCP servers.
Workflow
-
Clarify target briefly
- Identify the topic, output purpose, audience, deadline/currentness needs, and whether web search is allowed.
- If the task is understandable, proceed without asking extra questions.
-
Build an evidence map
- Read local/user materials first.
- Search the selected knowledge base when the user has selected KB materials or asks to use workspace knowledge.
- Use web search only when the user asks for current/public verification, the task requires facts not present in local materials, or the parent agent has already chosen web research.
- Keep search light: normally 2-5 focused queries; stop when 3-6 useful sources cover the question.
-
Assess source quality
- Prefer official, primary, recent, specific, and reproducible sources.
- Treat blogs, forums, social posts, AI-generated pages, and unsourced summaries as weak evidence unless the task is about public opinion.
- Mark outdated, inaccessible, paywalled, conflicting, or low-specificity sources.
-
Synthesize
- Extract claims, facts, examples, definitions, numbers, and caveats.
- Compare sources instead of merging conflicts silently.
- Explain what is well-supported, what is uncertain, and what still needs user confirmation.
-
Choose output
- Default: reply inline with a concise brief and source table.
- If the user asks to save, share, hand off to PPT/DOCX, or says "报告/简报/展示页/HTML/Markdown/文件", write files.
- For file output, generate Markdown first, then an HTML display page when useful.
Markdown Brief
Use this structure unless the parent agent needs a different format:
# [简报标题]
## 结论摘要
- [核心结论 1] [来源/资料标记]
- [核心结论 2] [来源/资料标记]
## 依据梳理
| 要点 | 依据 | 来源类型 | 可靠性 | 注意事项 |
| --- | --- | --- | --- | --- |
## 分析
## 不确定点与建议补充
## 来源表
| 编号 | 标题/文件 | 机构/作者 | 日期 | 类型 | 链接/路径 |
| --- | --- | --- | --- | --- | --- |
For education and learning tasks, add grade/subject/curriculum assumptions when relevant. For office tasks, add action-oriented recommendations and next steps.
HTML Display Page
Generate HTML when:
- the user explicitly asks for HTML, webpage, visualization, display, shareable brief, or preview;
- the parent task is preparing a presentation, teaching material, briefing page, or report artifact;
- a compact visual summary would materially help the user review the evidence.
HTML is a display layer, not the full research archive. Keep it concise and self-contained.
HTML requirements:
- Single
.html file, all CSS/JS inline.
- No CDN, external fonts, remote images, external scripts, trackers, or network requests.
- System font stack only.
- Responsive layout with no horizontal overflow on mobile.
- Include: title, 3-6 key findings, evidence matrix, limitations, and source table.
- Use simple CSS cards/tables; add inline SVG or CSS charts only when the data is real and sourced.
- Do not hide uncertainty behind polished visuals.
Suggested HTML sections:
Header: title, purpose, date, source scope
Key findings: 3-6 cards with source marks
Evidence matrix: claim / evidence / reliability / caveat
Implications or next steps
Limitations
Sources
Reliability Rules
- Do not fabricate URLs, document titles, page numbers, official wording, dates, authors, or statistics.
- If search is unavailable or GitHub/official sites are inaccessible, say so and continue with local materials, KB, or clearly marked general knowledge.
- If local documents and web sources conflict, surface the conflict and recommend which source should be treated as authoritative.
- If the task needs current facts but web search is unavailable, label the answer as "未联网核验".
- Avoid over-searching. For normal agents, this skill should stay lightweight.
Handoff Rules
- For PPT or lesson slides: hand the synthesized source notes to
pptx-deck-skill, html-ppt-skill, ai-animation-skill, or officecli-skills.
- For teacher-specific resource gathering, coordinate with
teaching-resource-research-skill when the task is about curriculum standards, teaching cases, public classroom resources, or educational media.
- For final review/exam tasks, hand concepts, source-backed scope, and uncertain exam claims back to
review-plan, knowledge-organize, practice-quiz, or exam-prep.
- For long, high-stakes research, escalate to
deep-researcher.
Final Response
When files are produced, list:
- Markdown path
- HTML path, if generated
- 2-4 sentence summary
- important limitations or sources that could not be verified
Do not include temporary paths.