research-and-writing

name

research-and-writing

description

Full-cycle research and writing skill for both Chinese thesis and English conference/journal papers. Covers: research idea incubation (literature landscape, GAP analysis, idea evaluation), research PRD planning, paper storyline design, experiment planning, section-by-section writing, translation, condensing/expanding, polishing, de-AI/humanization, architecture diagrams, experiment visualization, figure/table captions, experiment analysis, claim-evidence alignment, and adversarial self-review. Use when the user asks about: research ideas, literature survey, gap analysis, paper planning, thesis/paper writing, drafting, polishing, de-AI, logic checks, figure briefs, experiment design, claim-evidence checks, self-review, translation, condensing, expanding, architecture diagrams, experiment visualization, or any academic writing task.

Research Paper Writing Skill

Routing

This is the unified entry point for a complete research writing workflow. Identify the user's intent and route to the right section or reference accordingly.

Language Mode Selection (Mandatory for Writing & Review)

When the user's request involves writing or review, you must ask before proceeding:

"你当前的写作目标是？

1. 中文学位论文（毕业论文 / 学位论文）
2. 英文投稿论文（会议 / 期刊投稿）

请告诉我你的选择。"

Rules:

• Never guess the language mode from the user's input language.
• Once selected, persist for the session unless the user explicitly switches.
• Choice 1 → Apply Chinese Thesis conventions from references/chinese-thesis-style.md.
• Choice 2 → Apply English Paper conventions from references/english-paper-style.md.

Intent Routing Table

User Intent	Route To
idea, 研究方向, 选题, brainstorm, 文献调研, gap, literature survey	references/idea-incubation/ — see Idea Incubation
研究规划, 研究PRD, 技术路线, research plan, 可行性	references/research-prd/ — see Research PRD
storyline, paper story, 逻辑骨架, 大纲, outline	references/paper-storyline/ — see Paper Storyline
实验设计, 实验规划, ablation, experiment plan	references/experiment-planning/ — see Experiment Planning
翻译, translate, 中转英, 英转中	Writing Mode: Translation → references/paper-writing/translation.md
缩写, 扩写, condense, expand, shorten, lengthen	Writing Mode: Condensing/Expanding → references/paper-writing/condensing-expanding.md
写论文, draft, rewrite, abstract, introduction, method, experiments, related work, conclusion	Ask Language Mode → Section guides in references/paper-writing/
润色, polish, refine	Ask Language Mode → Writing Mode: Refine
去AI味, humanize, de-AI, renwei	Writing Mode: Humanize → references/refinement-modes.md
逻辑检查, logic check, consistency	Writing Mode: Logic Check
架构图, architecture diagram, framework figure	references/paper-writing/architecture-diagrams.md
实验绘图, visualization, chart recommendation	references/paper-writing/experiment-visualization.md
图标题, 表标题, caption, figure title, table title	references/paper-writing/figure-table-captions.md
实验分析, experiment analysis, data analysis	references/paper-writing/experiment-analysis.md
claim, evidence, 对齐, alignment	references/claim-evidence/ — see Claim-Evidence Alignment
审稿, review, 自查, 审核精修, defense, reviewer simulation	Ask Language Mode → references/review/

Global Drawing Convention

Scenario	Tool	Format
Flowcharts, architecture diagrams, logic maps	Mermaid	Embedded in Markdown
Precise technical figures, pipeline diagrams	DrawIO	.drawio file
Vector graphics requiring fine control	SVG	.svg file
❌ Forbidden	ASCII art / character drawings	Never use

Human-in-the-Loop Checkpoints

Pause and request user confirmation at these decision points:

• Idea incubation: GAP analysis results, Idea evaluation matrix, Idea Card finalization
• Research PRD: v1 concept doc confirmed before entering v2
• Paper storyline: Outline finalization
• Writing: Language mode selection
• Claim-evidence: Handling strategy for unsupported claims
• Review: Major issue fix proposals

Core Output Convention

• Default output is paste-ready finished content, not verbose explanations.
• For edits, return full revised text unless the user asks for diff-style output.
• Never fabricate citations, experiment data, or author/DOI information.
• When information is insufficient, mark [待补充] / [TO VERIFY] instead of guessing.
• Load only the references needed for the current task.
• Load examples on demand, not preloaded.

---

Writing Modes

Choose the smallest mode that fits the request. All modes apply to both Chinese and English unless noted.

1. Draft Mode

Use when the user asks for a new chapter, section, subsection, abstract, conclusion, or method description.

Rules:

• Write in the selected language mode (Chinese or English)
• Prefer full paragraphs over outline fragments
• Keep each paragraph focused on one main point
• Add figure notes or formulas only if useful
• Chinese: formal, specific, thesis-appropriate; load references/chinese-thesis-style.md
• English: clear, precise, conference/journal quality; load references/english-paper-style.md

2. Refine Mode

Use when the user already has a draft and wants it polished.

Rules:

• Preserve meaning, structure, and author intent
• Edit only where clarity, grammar, logic, or academic tone truly improves
• If the text is already solid, keep it close to the original
• Do not rewrite aggressively just to create visible change
• If the draft contains clear redundancy, repeated claims, pseudo-classification, or filler, structural reduction is preferred over cosmetic polishing

Load references/refinement-modes.md § Conservative Polish for detailed rules.

3. Humanize / De-AI Mode

Use when the user asks for "humanize", "de-AI", "去AI味", "renwei", or wants less machine-like prose.

Rules:

• Remove empty rhetoric and inflated significance claims
• Reduce mechanical patterns (repetitive connectors, forced symmetry, template phrasing)
• Preserve technical terms and domain meaning
• Never make the prose casual or chatty
• If machine feel comes from structure, delete or merge redundant material instead of rephrasing
• Em dashes: remove and rewrite naturally (applies to both languages)

Load references/refinement-modes.md § Humanize for full pattern list and examples.

Chinese-specific patterns: load references/chinese-thesis-style.md English-specific patterns: load references/english-paper-style.md

4. Translation Mode

Use when the user wants to translate between Chinese and English.

Load references/paper-writing/translation.md for:

• 中→英 (Chinese draft → English LaTeX): academic rewrite, proper escaping, tense rules
• 英→中 (English LaTeX → Chinese text): faithful direct translation for comprehension
• 中→中 (Chinese draft → formal Chinese): academic rewrite for Word
• Quality self-check protocols

5. Condensing / Expanding Mode

Use when the user wants to adjust text length without changing meaning.

Load references/paper-writing/condensing-expanding.md for:

• Condensing: reduce ~5-15 words by compressing syntax and removing fillers
• Expanding: add ~5-15 words by explicating implicit logic and enhancing connections

6. Logic-Check Mode

Use when the user asks for a logic check, consistency pass, or wants only real problems called out.

Rules:

• Use a high threshold; default assume the text has been through multiple rounds of editing
• Report only substantive issues
• Ignore cosmetic style preferences

Focus on:

• Contradictions
• Undefined terms or symbols
• Claim-evidence mismatch
• Missing transitions that block understanding
• Inconsistent naming of the same concept
• Duplicated claims across sections
• Pseudo-categories whose items are actually causes, effects, or conditions of one another
• Section-intro or section-ending sentences that only restate the heading
• Over-connected causal chains that weaken rather than strengthen the logic
• Literature review paragraphs that only summarize without extracting limitations
• Conclusion sentences whose basis is missing or weaker than the wording

7. Review-Pass Mode

Use when the user wants a chapter reviewed from an advisor, reviewer, or defense perspective.

Rules:

• Summarize the section's claimed contribution or purpose
• Separate major issues from minor improvements
• Explain which issues are fixable by revision and which are structural
• Keep the advice practical and chapter-specific
• Every conclusion sentence must have a basis: literature, experiment, or theoretical analysis

For full review-refine loop: load references/review/six-pass-review-loop.md For review output format: load references/review/review-output-template.md

8. Reviewer Simulation Mode

Use when the user wants the paper reviewed from a conference/journal reviewer perspective.

Rules:

• Read as a critical but fair reviewer for the target venue
• Assess: contribution significance, writing clarity, experimental strength, evaluation completeness, method design soundness
• Output: Summary, Strengths, Weaknesses (with specific evidence), Rating, Strategic advice
• Distinguish "fatal issues" from "fixable-in-revision issues"
• Score must reflect actual contribution level; do not default to harsh preset

Load references/review/five-dimension-review.md for the 5-dimension checklist.

9. Figure-Brief Mode

Use when the user needs help inserting or designing a figure.

Output:

• Insertion point
• Figure purpose
• Key visual elements
• Suggested caption
• Prose callout sentence
• Optional English figure title

For architecture diagram generation: load references/paper-writing/architecture-diagrams.md For experiment chart recommendations: load references/paper-writing/experiment-visualization.md For caption formatting: load references/paper-writing/figure-table-captions.md

10. Experiment Analysis Mode

Use when the user wants experiment results analyzed and written up.

Load references/paper-writing/experiment-analysis.md for:

• Data-driven analysis writing
• LaTeX \paragraph{} structure for experiment sections
• Truthfulness rules: all conclusions must come from data
• Output format with analysis paragraphs and Chinese back-translation

11. Formula Mode

Use when the user asks for formulas or wants formulas inserted into a section.

Output raw LaTeX only. No $, $$, \(, \), \[, \] delimiters. Keep notation consistent with the surrounding text. Add a short explanation in the user's language when useful.

---

Verification and Citation Rules

Apply sentence-level checking before finalizing any text (both Chinese and English).

For each sentence:

• Classify as factual, interpretive, methodological, or connective
• Re-check factual and literature-backed statements before keeping them
• Remove or soften claims that cannot be supported

Citation format:

• Chinese thesis: （Author, Year, DOI: xxx） (full-width parentheses)
• English paper: \cite{key} or (Author et al., Year) per venue style

Hard rules:

• Never invent author names, years, or DOI values
• When citations are missing or incomplete: proactively use runSubagent to search for and retrieve complete references (author names, years, DOI, publication venue). Do NOT leave placeholders like [TO ADD CITATION].
• The agent MUST execute literature search automatically when it encounters:
  • incomplete citations (missing DOI or publication year)
  • sentences with strong factual claims but no source nearby
  • any reference mentioned without full details
• Use search queries targeting arXiv, Google Scholar, IEEE Xplore, conference proceedings
• If a DOI is not verified after search, do not output a fake citation
• If support is missing after thorough search, use [待补充] / [TO VERIFY CITATION] as last resort
• If a sentence makes a strong claim but no source is available, rewrite conservatively or flag for verification
• Conclusion sentences must be backed by: literature, experiment, or explicit theoretical analysis
• This verification covers ALL sections, not only literature review

Load references/citation-verification.md for the full workflow.

---

Writing Convention Rules

Separate "de-AI" issues from "writing convention" issues. Convention rules are not mere style preferences.

Shared Conventions (Both Languages)

• First mention of an abbreviation → give full name, then abbreviate consistently
• Published works → prefer method name or paper title at first mention, not just "Author et al."
• Keep technical terms, abbreviations, and citation references stable across sections
• Do not treat convention rules as optional style suggestions

Chinese-Specific Conventions

• No spaces between CJK and Western letters/numbers: "ViT模型"、"LoRA微调"、"第3章"
• Use paired full-width quotation marks: ""和''
• Full-width Chinese punctuation: ，。；：
• First English term → full English name + Chinese translation, then abbreviation
• Do not use em dashes (——) in thesis prose; rewrite with colons, commas, or restructured clauses

Load references/chinese-thesis-style.md for the complete style guide.

English-Specific Conventions

• No contractions in academic prose (use "does not", not "doesn't")
• Avoid possessive forms for method/model names (use "the performance of X", not "X's performance")
• Present tense for methods and findings; past tense only for specific historical events
• Avoid AI-overused vocabulary: leverage, delve, tapestry, etc.
• No \textbf{} or \emph{} for emphasis in body text; use sentence structure
• Do not expand common domain abbreviations (keep LLM, not Large Language Models)
• Preserve existing LaTeX commands (\cite{}, \ref{}, \eg, \ie)

Load references/english-paper-style.md for the complete style guide.

---

Output Rules

Section Drafts

When drafting a section:

• Chinese: contiguous paragraphs, no empty blank lines between paragraphs, no Markdown formatting, suitable for direct paste into Word
• English: LaTeX-ready content with proper escaping (%, _, &)
• When any sentence needs a citation, AUTOMATICALLY search and retrieve the complete reference before finalizing. Use runSubagent if internal knowledge is incomplete.
• Run one final pass to check whether each factual sentence is supported, softened, or marked for verification

Refinement Passes

When refining existing text:

• Return the revised full text unless the user asks for diff-style edits
• Keep commentary short; mention only the most important changes

Figure Guidance

For each figure, provide:

• Insertion point, figure purpose, key visual elements, suggested caption, prose callout sentence
• If the figure does not exist yet, provide a creation brief

Formulas

Correct: L = \sum_{i=1}^{n}(y_i - \hat{y}_i)^2 Incorrect: $L = \sum_{i=1}^{n}(y_i - \hat{y}_i)^2$

---

Review-Refinement Loop (审核-精修循环)

Use when the user asks for "审核-精修循环", "审稿-精修", "review-refine loop", or "审核直到通过".

This is a composite workflow that runs 6 review passes, collects all issues, applies fixes, then re-reviews until no major issues remain. Works for both Chinese thesis and English papers.

Load references/review/six-pass-review-loop.md for the full 6-pass protocol.

Pass Summary

1. Reader Test: reading flow, undefined terms, cross-references, redundancy
2. De-AI / Humanize: mechanical patterns, template phrasing (load references/refinement-modes.md)
3. Writing Convention: terminology, abbreviation, citation format consistency
4. Citation & Fact Verification: sentence classification, citation support (load references/citation-verification.md)
5. Logic Check: contradictions, claim-evidence mismatch, naming consistency
6. Reviewer Pass: contribution claims, evidence backing, structural issues

Output Shape

[审核结果 - 第 N 轮] / [Review Result - Round N]

重大问题 / Major Issues:
1. ...

次要问题 / Minor Issues:
1. ...

已修复 / Fixed:
- ...

结论：审核通过 / 需要再修一轮
Conclusion: Passed / Needs another round

---

File-Oriented Workflow

If the user wants files created, write the final content to .txt files directly.

Suggested structure when none is provided:

• thesis/chapters/ch01-introduction.txt
• thesis/chapters/ch02-related-work.txt
• thesis/chapters/ch03-method.txt
• thesis/chapters/ch04-experiments.txt
• thesis/figures/figure-notes.txt

For finer granularity, split by section (e.g., 1.1.txt, 1.2.txt), then concatenate into a master draft (e.g., 第一章_总稿.txt).

Concatenation Rule

Use terminal commands (Get-Content in PowerShell or cat in bash) to concatenate files instead of reading each into context. After concatenation, use read_file and replace_string_in_file for edits.

---

Section Writing Guides

These references provide section-specific guidance. They apply to both Chinese and English papers. The section structure, logic flow, and argumentation principles are language-agnostic; apply the language-specific style conventions from references/chinese-thesis-style.md or references/english-paper-style.md to the final prose.

• references/paper-writing/abstract.md — Abstract: 3 templates (Challenge→Contribution/Insight)
• references/paper-writing/introduction.md — Introduction: backward reasoning + 4+3+4 template system
• references/paper-writing/related-work.md — Related Work: organization strategies
• references/paper-writing/method.md — Method: three-element module structure (Motivation→Design→Advantage)
• references/paper-writing/experiments.md — Experiments: 3 core questions + table rules
• references/paper-writing/conclusion.md — Conclusion
• references/paper-writing/does-my-writing-flow-source.md — Paragraph clarity checker
• references/paper-writing/paper-review.md — Self-review checklist

Examples Library

Read references/paper-writing/examples/index.md for the full example catalog. Load specific examples on demand:

• references/paper-writing/examples/abstract/ — 3 abstract template examples
• references/paper-writing/examples/introduction/ — 13 introduction examples (task intro + challenge + pipeline)
• references/paper-writing/examples/method/ — 9 method examples (three-element structure, module design)

Core Principles (Both Languages)

• One paragraph, one message.
• Reverse-outline verification: extract each paragraph's topic sentence; the sequence should form a coherent argument.
• Module description follows the triad: Motivation → Design → Technical Advantage.
• Tables use booktabs style (no vertical lines, minimal horizontal lines).
• Every major claim must be supported by experimental evidence (see claim-evidence alignment).

---

Idea Incubation

When the user asks about research ideas, topic selection, literature survey, or gap analysis, load and follow the references in references/idea-incubation/:

• references/idea-incubation/literature-landscape.md — Literature landscape search strategy and output template
• references/idea-incubation/gap-analysis.md — GAP identification methods and scoring rubric
• references/idea-incubation/idea-evaluation.md — Three-dimension evaluation framework (academic value / feasibility / resource fit)
• references/idea-incubation/idea-card-template.md — Final Idea Card output template

Idea Incubation Workflow

1. Anchor direction: Ask the user to describe their research interest
2. Literature landscape: Search and classify core papers by technical route
3. Contribution enumeration: Extract what/why-works/limitation per paper
4. GAP analysis: Identify unresolved challenges across routes
5. Idea seed generation: Propose 2-3 candidate ideas from GAPs
6. Idea evaluation: Score with three-dimension matrix
7. Human-in-the-loop: User reviews → feedback → iterate until confirmed
8. Idea Card finalization: Output structured Idea Card

---

Research PRD

When the user asks about research planning, technical roadmap, or feasibility, load and follow the references in references/research-prd/:

• references/research-prd/concept-doc-template.md — v1 research concept document template
• references/research-prd/landing-doc-template.md — v2 research landing document template
• references/research-prd/three-perspective.md — Three-dimension diagnostic detailed guide

Research PRD Workflow

1. Three-dimension diagnostic: Academic value → Technical feasibility → Resource constraints
2. v1 concept document: Direction alignment (user must confirm before v2)
3. v2 landing document: Executable technical plan + experiment design + timeline

---

Paper Storyline

When the user asks about paper story, outline, or logic skeleton, load and follow the references in references/paper-storyline/:

• references/paper-storyline/backward-forward-reasoning.md — Backward reasoning + forward narrative guide
• references/paper-storyline/introduction-logic-map.md — Introduction template catalog (Part A/B/C with multiple versions)
• references/paper-storyline/storyline-templates.md — 5 common paper type storyline skeletons

Chinese Thesis Chapter Arrangement (Pre-Step)

When the user is writing a Chinese degree thesis and asks about chapter arrangement, outline, or overall structure, apply the following rules before generating any outline:

1. Ask whether the user has reference theses: Instruct the user to obtain completed theses from their advisor or senior labmates in the same research group. These provide the most reliable signal for what chapter structure, depth, and length are expected by the specific institution and advisor.
2. Reference thesis is mandatory input: Do not generate a final chapter arrangement without the user confirming they have consulted (or chosen not to consult) at least one reference thesis. If they have one, ask them to describe the structure so the AI-generated outline can align.
3. AI assistance is secondary: The AI can propose a chapter skeleton based on the research topic and common conventions, but the reference thesis from the same group takes precedence when there is a conflict.
4. Human-in-the-loop checkpoint: After proposing a chapter arrangement, pause and ask the user to verify it against their reference thesis before proceeding to section-level writing.

Paper Storyline Workflow

1. Backward reasoning: Answer 4 key questions about contributions
2. Forward narrative: Build the story arc (Task→Challenge→Solution→Advantage)
3. Introduction logic map: Select template combination (A-VX + B-VX + C-VX)
4. Full-paper outline: Each section's core message + paragraph roles
5. Claim-evidence pre-mapping: Map each claim to its evidence source

---

Experiment Planning

When the user asks about experiment design, ablation planning, or experiment logs, load and follow the references in references/experiment-planning/:

• references/experiment-planning/experiment-design-matrix.md — Claim-Experiment mapping matrix template
• references/experiment-planning/ablation-checklist.md — Ablation derivation rules and template
• references/experiment-planning/experiment-log-template.md — Structured experiment recording template

Experiment Planning Workflow

1. Extract claims from storyline/outline
2. Design experiments: Map each claim to a validation experiment
3. Generate ablation checklist from pipeline modules
4. Prepare experiment log templates for structured recording
5. Track and update as a living document throughout research

---

Claim-Evidence Alignment

When the user asks about claim-evidence checks or alignment, load and follow the references in references/claim-evidence/:

• references/claim-evidence/claim-extraction.md — Rules for identifying claims in paper text
• references/claim-evidence/evidence-mapping.md — Criteria for evidence sufficiency and consistency
• references/claim-evidence/alignment-report-template.md — Output report template

Three Modes

• Build: Extract all claims from Abstract + Introduction, build mapping table
• Verify: Check each claim for sufficient evidence from experiments
• Fix: Soften unsupported claims or flag experiments to add

---

Review (5-Dimension + 6-Pass Loop)

When the user selects review mode, load references/review/:

• references/review/five-dimension-review.md — 5-dimension adversarial review checklist
• references/review/six-pass-review-loop.md — 6-pass review-refine loop (works for both Chinese thesis and English paper)
• references/review/review-output-template.md — Review output format templates (both languages)

5-Dimension Review (for conference/journal papers)

1. Contribution: New knowledge? Non-obvious novelty?
2. Writing Clarity: Reproducible? Consistent terms? One paragraph, one message?
3. Experimental Strength: Meaningful improvements over strong baselines?
4. Evaluation Completeness: All key ablations? Strong baselines included?
5. Method Design Soundness: Realistic setting? Robust without per-case tuning?

6-Pass Review Loop (for both Chinese thesis and English paper)

See Review-Refinement Loop above.

---

Shared Utilities

These cross-cutting references can be loaded by any module when needed:

• references/cross-section-consistency.md — Cross-section terminology, notation, abbreviation, and cross-reference consistency checks
• references/latex-engineering.md — LaTeX table/figure/formula formatting conventions and recommended packages
• references/output-format.md — Exact output shapes for all modes
• references/refinement-modes.md — Conservative polish, humanize/de-AI, logic check, and review pass detailed rules
• references/chinese-thesis-style.md — Chinese thesis tone, paragraphing, and anti-slop guidance
• references/english-paper-style.md — English paper tone, word choice, de-AI patterns, and LaTeX conventions
• references/citation-verification.md — Citation workflow and sentence-level fact checking