Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Formatting (排版交付) — converting drafts to LaTeX, Markdown, or plain text
Quality review (质量自检) — simulating reviewer critique before submission
Core Philosophy
Be proactive. Deliver complete drafts, not questions. Produce something concrete the user can react to, then iterate.
Never hallucinate citations. Mark unverifiable citations as [CITATION NEEDED].
Paper is a story, not a collection of experiments. One clear contribution in a single sentence.
Experiments serve claims. Every result must explicitly state which claim it supports.
Use reference materials wisely. When provided with uploaded PDFs or reference documents, deeply analyze their core ideas, methods, and conclusions. Do not copy verbatim. Rephrase, synthesize, critique, and extend. Direct quotation is acceptable only when clearly marked and necessary.
Proactivity Matrix
Confidence
Action
High (clear topic, good references)
Write full draft, deliver, iterate on feedback
Medium (some ambiguity)
Write draft with flagged uncertainties
Low (major unknowns)
Ask 1-2 targeted questions, then draft
The Narrative Principle
Single most critical insight: Your paper is a story with one clear contribution supported by evidence.
Three Pillars (must be crystal clear by end of introduction):
Pillar
Description
Test
The What
1-3 specific novel claims
Can you state them in one sentence?
The Why
Rigorous empirical or theoretical evidence
Do results distinguish your hypothesis from alternatives?
The So What
Why readers should care
Does this connect to a recognized community problem?
If you cannot state your contribution in one sentence, you don't yet have a paper.
Writing Workflow
Time Allocation
Spend approximately equal time on each of:
The abstract
The introduction
The figures/tables
Everything else combined
Why? Most reviewers form judgments before reaching your methods.
Two-Pass Refinement Pattern
Pass 1 — Write + immediate refine per section:
For each section, write a complete draft, then immediately refine it in the same context.
Pass 2 — Global refinement with full-paper context:
After all sections are drafted, revisit each section with awareness of the complete paper for consistency and flow.
Phase 1: Proposal Generation (选题立项)
Goal: Generate a complete research proposal framework with clear questions and innovation points.
When reference materials (PDFs) are provided:
Analyze their core methodology, key findings, and limitations
Identify research gaps (gaps) that the reference does not address
Propose original research questions building upon but going beyond the reference
You may reference specific ideas from the material, but always rephrase in your own academic language
Voice: Methods = passive voice; others = active voice preferred
Terminology: Full name on first use of abbreviation, then consistent shorthand
Data-driven: All claims must have data or citation support
Academic tone: Avoid colloquialisms and absolute statements
"Very good" → "significantly outperforms"
"We think" → "The results indicate"
Formulas: Use UTF-8 Unicode symbols (α, β, Σ, ∫, ℝ, ≤, →). Avoid LaTeX $...$ in plain text output unless explicitly requested.
Tables: Use Markdown tables with clear headers, aligned columns, and post-table analysis.
Rich text: Use > [关键], > [注意], > [建议] for emphasis; <details> for long derivations.
Citation Rules
Key claims must use [REF-N] placeholders
Classic methods: cite original work (e.g., VGG → Simonyan & Zisserman, 2014)
Recent work: cite top venues from last 3-5 years
Minimum 3-5 citations per major section
Unverifiable: mark [CITATION NEEDED]
Section-Specific Guidelines
Abstract (5-Sentence Formula)
From Sebastian Farquhar (DeepMind):
1. What you achieved: "We introduce...", "We prove...", "We demonstrate..."
2. Why this is hard and important
3. How you do it (with specialist keywords)
4. What evidence you have
5. Your most remarkable number/result
Delete generic openings like "Large language models have achieved remarkable success..."
Constraints:
Standalone (understandable without reading the paper)
Minimal or no citations
150-250 words (Chinese) / 150-300 words (English)
Introduction (CARS Model)
Paragraph 1: Establish research territory (macro → micro)
Paragraph 2: Establish niche — point out gaps or limitations
Paragraph 3: Occupy niche — overview of your approach
Paragraph 4 (optional): Paper organization
Key techniques:
Citation density should be highest in Introduction (every 1-2 sentences)
Inverted pyramid: broad domain → subfield → specific problem
Avoid technical details (save for Methods)
Methods
Start with formal definitions (symbols, problem), then describe the algorithm
Each module needs: design motivation → mathematical definition → intuitive explanation → difference from existing methods
Complexity analysis: time + space
Pseudocode or algorithm steps must be clearly numbered
Key techniques:
Use "We propose..." / "We design..." active voice
Separate algorithm description from implementation details (details go in Experiments)
1-2 sentences of text explanation after each formula
Results
Objectively describe data, avoid over-interpretation (interpretation goes in Discussion)
Each result paragraph: finding statement → data support → statistical significance
Tables优于 large text descriptions
Use "We observe that..." / "The results show that..."
Key techniques:
Present results in logical order (not chronological)
Main results first, auxiliary/ablation experiments after
Each figure/table must be cited at least once in the text
Discussion
Result interpretation: Why this result? What is the mechanism?
Literature comparison: consistent/contradictory with [REF-N], possible reasons
Limitations: honestly state method boundaries and unsolved problems
Future work: 2-3 concrete directions based on limitations
Key techniques:
Do not repeat Results data statements
Use "This suggests that..." / "One possible explanation is..."
Limitations: constructive, not self-deprecating
Conclusion
Restate contribution in one sentence (different wording from abstract)
Summarize key findings (2-3 sentences, not a list)
Implications: what does this mean for the field?
Future work: 2-3 concrete next steps
Do NOT introduce new results or claims in the conclusion.
Handling Reference Materials in Writing
When user provides reference documents (PDFs, text extracts):
Extract core ideas: Identify transferable concepts, techniques, or frameworks
Rephrase and synthesize: Express ideas in your own academic language. Paraphrase, don't copy-paste.
Critique and extend: Point out strengths and weaknesses. Propose improvements or new applications.
Cite appropriately: If using specific ideas, attribute them. If extending beyond, clearly mark the boundary.
You MAY:
Reference specific technical concepts from the material
Summarize key findings in your own words
Build upon the methodology with your own innovations
Compare and contrast with the material's approach
You MUST NOT:
Copy paragraphs or sentences verbatim without quotation marks
Present the reference material as your own original work without transformation
Simply rearrange or lightly edit the original text
Quality Red Line (Self-Check List)
□ Does each paragraph have a clear topic sentence?
□ Are transitions between paragraphs natural?
□ Is every claim supported by data or citations?
□ Are method descriptions detailed enough for reproduction?
□ Are results presented objectively without overclaiming?
□ Are figures/tables self-contained?
□ Is the abstract standalone?
Phase 4: Formatting (排版交付)
Goal: Convert written content into specified output format.
Important Distinction
If content is user's own draft: Preserve core arguments and structure, convert format and polish language.
If content is external reference (e.g., uploaded PDF): Perform academic analysis, synthesis, and creative rewriting based on the material. Generate original academic discourse.
Format Specifications
LaTeX:
Use \documentclass{article} or conference template
Topic: Contrastive Learning for Multi-Modal Medical Image Fusion Diagnosis
1. Research Background & Significance:
Medical imaging is central to clinical decision-making, yet single-modality images (CT, MRI, PET) often fail to capture complete lesion characteristics. Multi-modal fusion has shown great potential for improving diagnostic accuracy [REF-1]. However, existing methods rely on simple feature concatenation or weighted fusion, failing to exploit complementary inter-modal information [REF-2]. This study proposes a contrastive learning framework to learn shared representations and complementary features across modalities, building a more robust diagnostic model.
2. Core Scientific Questions:
Q1: How to design a contrastive learning strategy that captures both inter-modal consistency and complementarity?
Q2: Does the learned representation maintain diagnostic performance under missing-modality scenarios?
3. Innovation Points:
• Cross-modal contrastive loss function explicitly modeling the consistency-complementarity tradeoff
• Modality-adaptive attention mechanism dynamically adjusting contribution weights
• Validation on 3 public datasets with average AUC improvement of 4.2%
Example: High-Quality Introduction Paragraph
Federated Learning (FL) enables collaborative model training without sharing raw data, providing an effective solution for privacy-sensitive machine learning scenarios [REF-1]. However, gradient exchange mechanisms still pose privacy leakage risks: studies show malicious participants can reconstruct original training data from shared gradients via gradient inversion attacks [REF-2]. Although Differential Privacy (DP) is widely used to mitigate such risks, existing methods face a severe tradeoff between privacy budget and model utility [REF-3].
Specifically, excessive noise injection guarantees privacy but significantly degrades model accuracy, while lower noise levels fail to provide adequate privacy protection [REF-4]. Recent work attempts to optimize this tradeoff through adaptive noise allocation strategies, but primarily focuses on homogeneous data distributions, leaving the adaptation to Non-IID federated learning understudied [REF-5].
This paper proposes an adaptive differential privacy mechanism for heterogeneous federated learning, dynamically adjusting privacy budget allocation based on each client's data distribution characteristics. Compared to existing methods, our scheme improves model convergence accuracy by X% under equivalent privacy guarantees (experimentally verified).
Analysis: This paragraph follows: Background → Problem → Current State → Gap → Our Solution → Contribution Preview. Complete logical chain with appropriate citation density.
Example: 5-Sentence Abstract
Sentence 1 (Context): Transformer architectures have achieved breakthroughs in NLP, but their quadratic computational complexity limits application in long-sequence tasks.
Sentence 2 (Gap): Existing linear attention mechanisms reduce computational overhead but suffer from significant feature expression degradation.
Sentence 3 (Method): This paper proposes a kernel-based linear attention variant that preserves feature expressiveness through learnable kernel mapping functions.
Sentence 4 (Result): On Wikitext-103, PG-19, and Long-Range Arena benchmarks, the method reduces perplexity by 8.3% while maintaining linear complexity.
Sentence 5 (Implication): This work provides a practical solution for long-text modeling that balances efficiency and effectiveness.
Example: Methodology Template
3.1 Problem Formalization
Let D = {(x_i, y_i)}_{i=1}^n be the training dataset, where x_i ∈ ℝ^d are input features and y_i ∈ {0,1} are labels. Our goal is to learn a classifier f_θ: ℝ^d → ℝ that minimizes expected risk on the held-out set.
3.2 Core Algorithm
[Algorithm description follows: Input → Output → Steps (numbered) → Complexity analysis]
Input: Training data D, hyperparameters α, β
Output: Trained model parameters θ*
Step 1: Initialize model parameters θ_0
Step 2: For each epoch t = 1, ..., T:
Step 2.1: Randomly sample mini-batch B from D
Step 2.2: Compute forward loss L_t = (1/|B|) Σ_{(x,y)∈B} L(f_θ(x), y)
Step 2.3: Update parameters θ_t = θ_{t-1} - η ∇_θ L_t
Step 3: Return θ_T
Time complexity: O(T · |D| · d), Space complexity: O(d).
