gabrielmoreira/agent-skills-mirror/review-writing

Outline
  → [Phase 0: Validate & Revise outline]
  → [Phase 1..N: Per-section pipeline]
  → [Final: Assemble full review]
 
Each section is an independent unit of work:
  Search → Filter → Group → Write → Cite → Save to file
                                              ↓
                                    section_N.md (persisted immediately)
 
Final assembly reads all section files → cross-section dedup → unified numbering

## 大纲验证报告
 
### §1 [section title]
-检索命中: PubMed X篇, arXiv Y篇
-代表性论文:
  -"Paper Title A" (Journal, Year) — [关系:直接相关]
  -"Paper Title B" (Journal, Year) — [关系:方法论参考]
  -"Paper Title C" (Journal, Year) — [关系:最新进展]
-评估: ✅ 文献充足 / ⚠️ 偏少建议扩展 / 🔴 极少建议合并或调整
-建议: [specific suggestion if any]
 
### §2 ...
 
### 整体建议
-建议新增: [topic] — 检索发现大量文献但大纲未覆盖
-建议合并: §X 和 §Y 文献高度重叠
-建议拆分: §Z 文献过于丰富，建议拆为两节

From the following N search results, select the most relevant papers
for writing a review section about [section topic].
 
SELECTION TARGETS:
- Biomedical sections: select 10-15 papers
- CS/AI sections: select 15-20 papers
 
SELECTION CRITERIA (in priority order):
1. RELEVANCE to the specific section topic
2. IMPACT: prefer high-impact journals and highly-cited works
3. RECENCY: prefer 2022-2026, but include seminal older papers
4. DIVERSITY: cover different sub-aspects, not just the most popular finding
5. BALANCE: include both supporting evidence and contrasting viewpoints
 
For each selected paper, assign ONE role:
- FOUNDATIONAL: establishes the field/concept
- EVIDENCE: provides key experimental/computational evidence
- METHOD: introduces an important method/tool
- COMPARISON: enables comparison between approaches
- GAP: identifies limitations or open problems
- MILESTONE: landmark paper (e.g., AlphaFold, GPT-4)
 
Output format:
[search_index] [ROLE] — one-sentence reason for inclusion

Group A: "Historical development / Milestones"
Group B: "Current mainstream approaches / Consensus"
Group C: "Emerging advances / Recent breakthroughs"
Group D: "Methodological comparisons"
Group E: "Limitations, controversies, and open problems"

Write a review section about [section topic] using the provided literature.
 
WRITING RULES:
1. ACADEMIC CHINESE PROSE (学术中文). Scientific terms keep English abbreviations
   on first mention: e.g., 检索增强生成（Retrieval-Augmented Generation, RAG）.
   Subsequent uses can use abbreviation directly.
 
2. NARRATIVE, NOT LIST. Write flowing paragraphs with logical transitions.
   ❌ "A研究了X[1]。B研究了Y[2]。C研究了Z[3]。"
   ✅ "多项研究从不同角度探讨了这一问题。A等[1]首先通过...揭示了...；
      在此基础上，B等[2]进一步...；然而，C等[3]的研究指出..."
 
3. CITE BY INDEX. Reference papers using [N] where N is the paper's index in
   the provided source list. Every factual claim MUST have at least one citation.
 
4. CRITICAL ANALYSIS. Don't just summarize — compare, contrast, evaluate.
   Point out methodological differences, conflicting findings, remaining gaps.
   ❌ "取得了重要进展"、"具有广阔前景"
   ✅ "将检索精度从 70.1% 提升至 80.7%"、"覆盖了 338 个数据库"
 
5. SECTION STRUCTURE:
   a. 开门点题（1-2句）：本节综述什么主题，为什么重要
   b. 发展脉络（2-3段）：按时间或逻辑组织
   c. 现状分析（1-2段）：主流方法/共识/争议
   d. 批判性评价（1段）：现有工作的局限和不足
   e. 收束引出（1-2句）：指向下一节或研究空白
 
6. SECTION TRANSITION: The FIRST sentence of this section must logically connect
   to the LAST sentence of the previous section. The LAST sentence must set up
   the next section's topic. [Agent: verify this after writing.]
 
7. LENGTH: 1,500-2,500 Chinese characters per section.
 
8. NO FABRICATION. Only cite papers from the provided source list.
   If a fact lacks source support, write "据报道" without citation rather than
   fabricating one. NEVER invent PMIDs, DOIs, or author names.
 
9. JOURNAL NAMES: Use FULL journal names (Nature Medicine, not Nat Med).
   Keep this consistent across all sections.
 
SOURCE LIST:
[paste filtered papers with index, title, authors, year, abstract]

# Count paragraphs and citations
paragraphs = [p for p in section_text.split("\n\n") if p.strip()]
for i, p in enumerate(paragraphs):
    cite_count = len(re.findall(r"\[\d+\]", p))
    if cite_count == 0 and len(p) > 100:
        print(f"WARNING: Paragraph {i+1} has no citations ({len(p)} chars)")

output_dir/
  section_1_[short_name].md
  section_2_[short_name].md
  ...
  section_N_[short_name].md
  _metadata.json

# [Section Number] [Section Title]
 
[Section text with [N] citations]
 
---
## 本节参考文献（临时编号）
 
[1] Author, et al. Title[J]. Journal, Year, Vol(Issue): Pages. DOI: xxx.
[2] ...
 
---
<!-- metadata
section_index: 1
cited_papers: [
  {"local_index": 1, "pmid": "12345678", "doi": "10.1234/xxx", "title": "...", "source": "pubmed"},
  {"local_index": 2, "pmid": "", "doi": "", "arxiv_id": "2210.03629", "title": "...", "source": "arxiv"},
  ...
]
search_queries: ["query1", "query2", ...]
candidate_count: 25
cited_count: 14
-->

{
  "outline_file": "/path/to/综述大纲.md",
  "output_dir": "/path/to/综述输出/",
  "sections_completed": [1, 2, 3],
  "sections_total": 8,
  "all_cited_papers": [
    {"pmid": "12345678", "doi": "...", "title": "...", "first_cited_in_section": 1},
    ...
  ],
  "total_unique_references": 45,
  "seed_reference_pool": [...],
  "last_updated": "2026-02-27T20:30:00"
}

import re, json
def assemble_review(section_files, output_path):
    """Assemble all sections into final review with unified GB/T 7714 numbering."""
    global_refs = []
    paper_to_global = {}
    global_num = 1
    full_text_parts = []
    for sf in section_files:
        with open(sf) as f:
            content = f.read()
        # Split text from metadata
        text_part = content.split("---\n## 本节参考文献")[0]
        # Load cited papers from metadata comment
        meta_match = re.search(r'<!-- metadata\n(.*?)\n-->', content, re.DOTALL)
        local_papers = []
        if meta_match:
            meta_text = meta_match.group(1)
            cp_match = re.search(r'cited_papers: (\[.*?\])', meta_text, re.DOTALL)
            if cp_match:
                local_papers = json.loads(cp_match.group(1))
        if not local_papers:
            full_text_parts.append(text_part)
            continue
        paper_by_local = {p["local_index"]: p for p in local_papers}
        def remap(m):
            nonlocal global_num
            local_idx = int(m.group(1))
            paper = paper_by_local.get(local_idx)
            if not paper:
                return ""
            key = (paper.get("doi") or paper.get("pmid") or
                   paper.get("title", "").lower())
            if key not in paper_to_global:
                paper_to_global[key] = global_num
                global_refs.append(paper)
                global_num += 1
            return f"[{paper_to_global[key]}]"
        remapped = re.sub(r"\[(\d+)\]", remap, text_part)
        full_text_parts.append(remapped)