AgentSkillsCN

deep_research

系统化的研究工作流,助力证据整合与报告撰写。

SKILL.md
--- frontmatter
name: deep_research
description: Systematic research workflow for evidence synthesis & reporting

Deep Research

Repeatable workflow for complex questions requiring evidence synthesis & structured reporting.

Purpose

  • Scoped, multi-phase research
  • Gather & evaluate diverse sources w/ quality checks
  • Synthesize into themes, patterns, gaps, recommendations
  • Traceable citations

Core Principles

  1. Scope First: Define questions, criteria, constraints before search
  2. Source Diversity: Multiple independent sources, prefer primary
  3. Evidence Grading: Record recency, authority, confidence
  4. Synthesis > Summary: Themes, patterns, gaps
  5. Reproducibility: Track inputs, queries, decisions

Rules

Scope Definition

Capture main question, subquestions, success criteria, constraints, deliverables before search.

See doc/research_plan_template.md.

Source Collection

Collect 10-15+ sources across categories (official docs, academic, benchmarks, practitioner reports). Record author, date, link.

Source Evaluation

Score each for currency, relevance, authority, accuracy, purpose. Summarize confidence.

See doc/source_evaluation.md (CRAAP framework).

Deep Dive

Prioritize where importance & uncertainty high. Use citation chaining & triangulation.

Synthesis

Group by theme, identify patterns, capture gaps w/ confidence levels.

See doc/findings_template.md.

Reporting

Executive summary, methodology, findings, comparisons, recommendations, risks, references.

Use doc/final_report_template.md.

Automation (Optional)

Scripts = helpers, not substitutes. Doc assumptions & inputs.

Decision Framework

code
High impact, ambiguous, controversial:
  → Full deep research workflow

Well-known, low impact:
  → Lightweight scan, cite 2-3 sources

Time constrained:
  → Full workflow but reduced depth, doc limitations

Examples

Tech Evaluation

Compare Next.js vs Remix for SSR → structured report w/ scope, benchmarks, ecosystem, decision matrix, recommendations.

Market Research

AI coding assistant market → report w/ market signals, vendor positioning, trend analysis, confidence levels, gaps.

Edge Cases

  • Limited sources: doc gap, lower confidence, avoid strong recs
  • Conflicting evidence: present competing views, explain weighting
  • Paywalled: use abstracts/secondary, note limitations
  • Tight deadlines: reduce depth, keep methodology & limitations explicit

Related: thinking_framework, documentation, security_experts

See COMMON.md.

Tools: bin/search_and_analyze.py, bin/synthesize.py, bin/report_generator.py

Refs: CRAAP, PRISMA (2026-01-26)