AgentSkillsCN

deep-research

具备深度且经过严格溯源验证的研究能力,通过网络搜索与浏览,仅从官方权威渠道获取高可信度的事实,并以结构化、详尽的 Markdown 报告形式呈现(不提供摘要)。

SKILL.md
--- frontmatter
name: deep-research
description: Deep, source-verified research skill that uses web search/browse to gather only high-confidence facts from official and authoritative sources, and outputs a structured, detailed Markdown report (no summaries).
short-description: High-confidence research from official sources only
category: Information Gathering
tags:
  - research
  - verification
  - citations
  - official-sources
version: 1.0.0
author: Mihai Adrian
license: MIT
compatibility:
  - codex: "5.2+"

DeepResearch Skill

What This Skill Does

Produces a detailed Markdown research report based only on confirmed, high-confidence facts from official or authoritative sources. It always uses web search/browse tools and excludes summaries to avoid loss of context.

When To Use

  • The user asks for deep research or documentation on a topic.
  • The user requires official/authoritative sources and strong verification.
  • The result must be a full Markdown report (not a short summary).

Inputs To Clarify (Ask If Missing)

  • Topic and scope (what is in-scope vs out-of-scope)
  • Timeframe (e.g., 2023-2026) and region/jurisdiction
  • Depth expectations (number of sections, granularity)
  • Output path and filename (if not provided, choose a default)

Tool Requirement (Mandatory)

You MUST use the web search/browse tool (web.run) for this skill. Do not proceed without online verification.

Source Policy (Strict)

Accept ONLY authoritative sources such as:

  • Government or regulator sites (e.g., .gov, official EU sites)
  • Standards bodies (ISO, IEEE, W3C, etc.)
  • Peer-reviewed journals or reputable academic publishers
  • Official vendor documentation or official announcements
  • Audited datasets published by recognized institutions

Reject or treat as non-authoritative:

  • Personal blogs, social media, forums, marketing-only pages
  • Secondary summaries that lack primary citations
  • Undated or anonymous sources

If primary sources are unavailable for a claim, do not include the claim.

Confidence Model (Use This Exactly)

Define independent sources as separate organizations (not the same publisher syndicating content).

Confidence tiers:

  • 100% (Confirmed-3+): 3+ independent authoritative sources OR 2 authoritative + 1 peer-reviewed/standard, all consistent
  • 95% (Confirmed-2): 2 independent authoritative sources, consistent
  • 90% (Confirmed-1+2): 1 authoritative source + 2 reputable secondary sources that cite primary data, consistent

Only include findings with confidence >= 90%. If a topic area lacks such evidence, list it as a Research Gap without stating unconfirmed claims.

Research Workflow

  1. Scope and Decompose

    • Break topic into 3-7 research areas.
    • For each area, list the exact questions to answer.
  2. Source Discovery (via web.run)

    • Search for primary, official sources first.
    • Prefer original publications and authoritative domains.
  3. Source Validation

    • Verify authoring org, publication date, and originality.
    • Confirm independence among sources.
  4. Evidence Extraction

    • Extract facts, numbers, definitions, and formal statements.
    • Paraphrase and cite; do not rely on long quotes.
  5. Corroboration and Scoring

    • Group equivalent claims across sources.
    • Assign confidence per the model above.
    • Exclude anything below 90%.
  6. Report Generation

    • Write a structured Markdown report with full citations.
    • No summary section. Keep full context.

Output Requirements (Markdown File Only)

  • Always produce a Markdown file on disk.
  • If the user did not specify a path, use:
    • ./docs/research/ if it exists, else
    • ./research/, else
    • current working directory.
  • Filename format: DeepResearch_<topic_slug>_<YYYY-MM-DD>.md

Report Structure (Mandatory)

  1. Title
  2. Scope and Methodology
  3. Source Criteria and Confidence Model
  4. Detailed Findings (grouped by research area)
    • Each finding must include:
      • Statement
      • Evidence table (sources, dates, publisher)
      • Confidence tier and rationale
  5. Data Tables / Figures (if applicable)
  6. Research Gaps (no unconfirmed claims)
  7. Full Source Register (complete citations)

Example Invocation

User: "Run DeepResearch on EU AI Act enforcement timeline and obligations, 2024-2026, EU only."

Output: ./docs/research/DeepResearch_eu_ai_act_2026-01-13.md

Common Pitfalls

  • Using secondary summaries without checking primary sources
  • Including claims with only one non-authoritative source
  • Writing an executive summary (not allowed)
  • Mixing opinion with verified facts