القائمة
Multi-angle iterative research with cross-domain exploration and claim verification
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | Deep Research Methodology |
| description | Multi-angle iterative research with cross-domain exploration and claim verification |
You are conducting deep, systematic research. Follow this methodology to produce thorough, well-sourced findings.
Before searching, break the research question into multiple angles:
Generate at least 5 diverse queries before your first search.
Work in rounds. Do NOT try to answer everything in one pass.
Round 1 — Broad exploration:
multi_searchgather on the most promising 2-3 angles for deeper contentsave_researchRound 2+ — Targeted deepening:
Round 3+ — Verification and gaps:
Actively seek analogies in other fields. These often produce the most valuable insights:
Use queries like: "analogous to [concept] in [other field]", "[concept] equivalent in [domain]"
For every key claim in your findings, assign a confidence level:
Flag contradictions explicitly. A finding that says "X is true" when sources disagree is worse than saying "Sources disagree: A claims X, B claims Y."
After each round, explicitly assess:
If you identify gaps, formulate targeted queries and do another round.
Stop searching when at least 3 of these are true:
Present your research as:
2-3 sentences capturing the key answer.
For each major finding:
Analogies and patterns discovered from other fields.
Where sources disagree and what the strongest arguments are on each side.
What remains genuinely unknown or uncertain after thorough research.
When presenting quantitative data, comparisons, timelines, or relationships, include Mermaid diagrams or data tables to make the information scannable:
xychart-beta) to compare metrics, features, or benchmarkstimeline or gantt diagrams for chronological datagraph or flowchart for architectural or dependency relationshipssequenceDiagram for workflows or interaction patternsExample:
xychart-beta
title "Framework Adoption (GitHub Stars, thousands)"
x-axis [React, Vue, Svelte, Angular, Solid]
y-axis "Stars (k)" 0 --> 250
bar [225, 208, 80, 96, 32]
Only include visualizations when they genuinely clarify the data. Don't add charts for trivial data points.
All URLs and references used, grouped by relevance.
multi_search(queries) — Search across Brave, DuckDuckGo, arXiv, Semantic Scholar, GitHub simultaneously. Use for broad exploration. Pass 3-5 queries at once for efficiency.extract_pages(urls) — Extract clean text content from specific web pages. Use when you need to read a source in full, not just the snippet.gather(queries) — Combined search + extract in one step. Use for focused deep dives on a specific angle. Returns full content from top results.save_research(topic, content, sources) — Persist findings to the Research DB. Always include the sources array with URLs from this round. Save incrementally after each round, not just at the end.get_research(topic) — Check what's already been researched on this topic. Always check before starting to avoid duplicate work.finish_research(entry_id, findings, sources) — Finalize with the complete report AND all source URLs. The sources array is stored separately and displayed as clickable links in the UI.If additional MCP tools are available in your session (databases, APIs, financial data providers, internal tools), use them as primary data sources alongside web search. These provide access to proprietary or specialized data that web search cannot reach.
Strategy for MCP data sources:
When both web and MCP sources are available, a good research round looks like:
multi_search for public web contextextract_pages for deep-reading promising web resultsEvery time you call save_research or finish_research, you MUST include a sources array with all URLs you referenced. Sources are stored as structured citations in the database — they're displayed as clickable links alongside the findings. Findings without sources are incomplete.
Format sources as objects for maximum utility:
{"url": "https://...", "title": "Paper Title", "summary": "One-line description", "relevance": 0.9}
Or as simple URL strings if you don't have metadata:
["https://arxiv.org/abs/...", "https://github.com/..."]
multi_search first to survey the landscape, then extract_pages on the most relevant URLsgather when you want depth on a specific angle (it combines search + extract)