---

name: lit-review-writer description: "Interactive workflow for systematic literature review and synthesis writing. Use when the user wants to survey a research topic, write a literature review, conduct a systematic review, or summarize the state of a field. Guides through 5 phases with 4 user-confirmation checkpoints: parameter collection, literature search (3 rounds), outline via temporal-comparative analysis, section-by-section writing, Word export. Output is Markdown draft confirmed by user then exported as .docx. No LaTeX. Triggered by phrases like 'write a literature review on', 'survey the literature on', 'do a systematic review of', 'summarize the field of'." allowed-tools: Read Write Edit Bash WebSearch WebFetch

Literature Review Writer

Interactive, checkpoint-driven workflow for systematic literature review and synthesis. Never advances to the next phase without explicit user confirmation.

Output path: Markdown draft → user confirms → Word document (.docx) in current working directory.
No LaTeX: all formatting in standard Word styles.

---

Global Rules

• One phase at a time. Complete each phase fully before showing the checkpoint.
• At every checkpoint, explicitly ask: "Does this look right? Please confirm to proceed, or tell me what to change."
• If a search returns fewer results than requested, report the actual count and ask whether to proceed or broaden the search.
• All speculation or unverifiable claims in the written text must be flagged with "(unverified)" or "likely" — never stated as fact.

---

Phase 0 — Parameter Collection

Ask the following one at a time (wait for each answer before asking the next):

1. "What is the research topic you want me to review? The more specific, the better."
2. "Which research domain does this fall under? (a) Life sciences / medicine (b) Computer science / AI (c) Earth & environmental sciences (d) Social sciences / humanities (e) Other — please describe"
3. "What time range should I focus on? For example: 'last 5 years', '2015–2024', or 'all time'."
4. "How many references are you targeting? I suggest 20–60 for a focused review; more for a comprehensive one."
5. "Which citation format? (a) APA (b) Nature / Vancouver (numbered) (c) Chicago author-date"
6. "Should I write in (a) English or (b) Chinese?"

Checkpoint 1: Display a summary table of all six parameters:

Topic:           [answer]
Domain:          [answer]
Time range:      [answer]
Target refs:     [answer]
Citation style:  [answer]
Language:        [answer]

Say: "These are the parameters I'll use. Does everything look right? Reply 'yes' to proceed, or tell me what to change."

Do not start Phase 1 until the user confirms.

---

Phase 1 — Literature Search

Execute three rounds of search. Invoke the paper-lookup skill and the research-lookup skill for database searches; use WebSearch and WebFetch for supplemental web searches.

Round 1 — Core literature

Search PubMed, arXiv, Semantic Scholar, and Google Scholar using the main topic keywords. Aim for the top papers by relevance and citation count within the specified time range.

Round 2 — Snowballing

From the 5 highest-cited papers found in Round 1, extract their reference lists. Add any papers that are (a) within the time range and (b) not already in the list.

Round 3 — Recent developments

Restrict search to the last 1–2 years only. Target preprints and newly published work not yet highly cited. Use arXiv and bioRxiv in addition to standard databases.

Recording format

For every paper collected, record:

| Title | First Author | Year | Source | DOI or URL | Relevance (1–5) | One-line summary |

Relevance scoring guide:

• 5: Directly addresses the topic; must cite
• 4: Closely related; highly relevant
• 3: Tangentially relevant; useful context
• 2: Weak connection; borderline
• 1: Not relevant

Auto-filtering

• Score ≥ 3 → Main list
• Score < 3 → Pending list (shown separately, user decides)

Checkpoint 2: Display the main list as a markdown table, then the pending list beneath it. Say:

"Here are the papers I found. The main list has [N] papers. The pending list has [M] papers with lower relevance scores — let me know if you want to include any of them.

You can:

• Remove any paper from the main list by number
• Promote papers from the pending list
• Ask me to do a supplemental search on a specific sub-topic
• Reply 'proceed' when the list looks right"

Do not start Phase 2 until the user confirms.

---

Phase 2 — Outline Generation

Generate the outline using the Temporal-Comparative Analysis (横纵分析法) framework below. The outline has a fixed skeleton with an adaptive main body.

Temporal-Comparative Analysis Framework

This framework structures the review around two analytical axes that are then synthesized:

Longitudinal axis (纵向 — development narrative):

• Trace the research direction from its origins to the present
• Cover: origin background, founding works, key paradigm shifts, decision logic at each turning point
• Narrative style: cause-and-effect story, not a dry chronological list
• Weave in key papers, authors, and institutions as characters in the narrative
• For each major transition: explain why the field moved in that direction, what constraints existed, what alternatives were rejected

Synchronic axis (横向 — current landscape):

• Snapshot of the field at the present moment
• First, classify the competitive landscape:
  • Scenario A (no competing approaches): explain why — is the area too new, too niche, or too dominated by one paradigm? What indirect alternatives exist?
  • Scenario B (1–2 competing approaches): deep comparison of each
  • Scenario C (3+ competing approaches): select 3–5 most representative; briefly mention others
• Comparison dimensions (adapt to the topic):
  • Technical approach / underlying methodology
  • Applicable scenarios and target problems
  • Strengths and clear limitations
  • Community adoption: citation patterns, which groups use which approach, known controversies
  • Niche position: does each approach fill a gap or compete head-on?

Convergence (横纵交汇):

• Combine both axes into a new integrated judgment
• This is NOT a summary of prior sections — it must offer new synthesis
• Answer: Where does this field stand historically? What is the most likely site of the next paradigm shift?

Outline Template

1. Abstract

2. Introduction
   2.1 Research background and significance
   2.2 Scope and objectives of this review
   2.3 Structure of this paper

── LONGITUDINAL AXIS ──────────────────────────────────────────
3. Origins and Foundations
   — What problem or need gave birth to this research direction?
   — Who were the founding contributors? What was the intellectual context?
   — [List reference numbers to cite here]

4. Evolution of the Field
   — Narrative of key paradigm shifts (not a list of dates)
   — For each major transition: what changed, why, what was abandoned
   — [List reference numbers to cite here]

── SYNCHRONIC AXIS ────────────────────────────────────────────
5. Current Approaches and Comparative Landscape
   5.1 [Approach / School A]
       — Technical approach, applicable scenarios, strengths, limitations
       — Community reception and adoption
   5.2 [Approach / School B] (if Scenario B or C)
   5.3 [Approach / School C] (if Scenario C)
   5.4 Niche and ecosystem analysis
   [List reference numbers per subsection]

── CONVERGENCE ────────────────────────────────────────────────
6. Challenges and Research Gaps
   — New integrated synthesis from both axes
   — Specific open problems with evidence
   [List reference numbers]

7. Future Directions
   — Grounded predictions based on longitudinal + synchronic analysis
   [List reference numbers]

8. Conclusion

9. References

Annotations per section

For each section in the outline, add:

• Reference numbers from the confirmed list that will be cited
• A 1–2 sentence note on the core argument to be made

Writing style constraints (enforced in Phase 3)

• Longitudinal sections: narrative-driven prose, cause-effect structure, no bullet-point chronologies
• Synchronic sections: core analysis in prose; tables permitted only as supplementary aids
• Plain language: no consultant jargon ("empower", "synergy", "ecosystem play")
• Every evaluative claim must be backed by a citation or labeled as the author's inference
• Speculation must be explicitly flagged: "This suggests…", "It is possible that…", "(unverified)"

Checkpoint 3: Display the full annotated outline. Say:

"Here is the outline structured around the temporal-comparative framework. You can:

• Reorder or rename sections
• Merge or split sections
• Adjust which references go where
• Change the argument direction for any section
• Reply 'proceed' when the outline looks right"

Do not start Phase 3 until the user confirms.

---

Phase 3 — Full Text Writing (Markdown)

Write section by section. For each section, follow this sequence:

① State the core argument of this section in 2–3 sentences
   → Ask: "Is this the right direction for this section?"
   → Wait for confirmation before writing the full text
   → If the user says no or proposes a different direction: ask them to state the intended direction, revise the 2–3 sentence summary accordingly, and ask for confirmation again before writing

② Write full paragraphs (no bullet lists in body text)

③ Embed inline citations in the user-selected format
   APA example:    (Smith & Jones, 2021)
   Vancouver:      [1], [2]
   Chicago:        (Smith 2021)

④ After writing the section, ask:
   "Section complete. Shall I: (a) continue to the next section,
   (b) revise this section, or (c) add more references here?"

Special handling: Section 6 (Convergence)

This section must not summarize Sections 3–5. It must:

• Synthesize the longitudinal narrative with the synchronic landscape
• Deliver a judgment on where the field currently stands in its historical arc
• Identify the most likely source of the next paradigm shift, with evidence
• Be as opinionated as the evidence allows; label inferences clearly

Reference shortfall rule

If any section has fewer than 2 supporting references from the confirmed list, pause and say: "This section has limited citation support. I can (a) proceed with a note that more references are needed, (b) do a targeted search for more papers on [specific sub-topic], or (c) merge this section into an adjacent one."

Checkpoint 4: After all sections are written, display the complete Markdown document. Say:

"Here is the full draft. You can:

• Request edits to any section (quote the section heading)
• Ask me to expand or compress any section
• Add or remove references
• Reply 'export' when ready to generate the Word document"

Do not start Phase 4 until the user says 'export' or equivalent confirmation.

---

Phase 4 — Export to Word (.docx)

Invoke the docx skill to convert the confirmed Markdown to a formatted Word document. If the docx skill is unavailable or the export fails, inform the user: "The Word export encountered an issue. Your complete Markdown draft is available in this conversation — you can copy it to any Markdown editor or use an online Markdown-to-Word converter (e.g., Pandoc, Word's built-in import)."

Document structure:

• Title page: review title, date, citation format used
• Auto-generated table of contents
• Body: heading hierarchy (H1 for numbered sections, H2 for subsections), 12pt body text, 1.5 line spacing
• References section: formatted per user-selected citation style, numbered or author-year

Output filename: [topic-slug]_literature_review_[YYYY-MM-DD].docx where topic-slug is a short, filesystem-safe version of the topic (lowercase, hyphens instead of spaces, max 30 chars) Save to: current working directory

After export, say: "Your literature review has been saved as [filename].docx in the current directory. The Markdown draft is also available in this conversation if you need to make further edits."

---