Scientific Research Helper
NguyenAn2812/scientific-research-helper
Assistente completo para pesquisa acadêmica, teses e papers científicos
Ver no GitHubSkill para Claude que auxilia estudantes em todas as etapas da pesquisa acadêmica, desde brainstorming até verificação de outputs estatísticos. Analisa documentos, fornece feedback estruturado, sugere melhorias estruturais, validação de citações e aprimoramento da linguagem acadêmica.
Instalação
npx convez-ai NguyenAn2812/scientific-research-helper1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise estruturada de teses, dissertações e papers científicos
- Verificação de consistência entre dados, código e resultados reportados
- Pesquisa bibliográfica e detecção de saturação em tópicos acadêmicos
- Validação de citações e formatação de referências
Veredito

Skill bem-estruturada e útil para o contexto acadêmico brasileiro. Cobre ciclo completo de pesquisa com funcionalidades práticas (análise de arquivos, verificação de código/dados, busca bibliográfica). Limitação: depende de acesso a APIs externas (Google Scholar, Semantic Scholar) e pode ter variação na qualidade de feedback em disciplinas muito especializadas.
README
A tradução para o português deste README ainda está na fila.
Scientific Research Helper — Claude Skill
A Claude skill that guides students through every stage of academic research — from brainstorming a topic to verifying statistical outputs.
What This Skill Does
Scientific Research Helper turns Claude into a dedicated academic research companion for undergraduate and graduate students across all disciplines. Instead of generic AI responses, you get a structured, stage-aware assistant that knows where you are in the research process and what you actually need next.
Whether you're staring at a blank page trying to pick a topic, stuck on your methodology section, or getting feedback on a draft thesis — this skill has a workflow for it.
Key Features
File Review — Upload & Get Structured Feedback
Drop in a PDF, DOCX, or paste your draft. The skill reads it end-to-end, identifies the document type (thesis, IMRAD paper, research proposal, internship report, etc.), and gives you structured feedback across three dimensions:
- Consistent — strengths worth keeping
- Needs review — ambiguous or underdeveloped areas
- Clear contradictions — internal inconsistencies flagged with specific quotes
Writing Plan & Asset Inventory
Tell the skill your topic, share your evidence files, and it generates a complete inventory of everything your paper needs: a personalized section outline, a full table list, and a figure list — each with instructions on how to extract or produce the asset from your actual data files.
Literature Search & Overlap Check
Before you invest weeks into a topic, the skill searches Google Scholar, Semantic Scholar, and ResearchGate to assess how saturated the research space is — and suggests angles for original contribution. It can also spot-check citations in your bibliography for accuracy.
Data & Code Verification
Paste your Python, R, SPSS, Stata, or MATLAB files alongside your report. The skill cross-references the numbers in your text against your actual analysis output and flags any discrepancies — before your supervisor does.
Supported Research Stages
| Stage | What the Skill Helps With |
|---|---|
| 1 — Brainstorming | Topic ideas, research questions, feasibility check |
| 2 — Proposal | Objectives, hypotheses, scope, theoretical framework |
| 3 — Literature Review | Search strategy, clustering themes, writing the review |
| 4 — Methodology | Quantitative / qualitative / mixed methods guidance |
| 5 — Review & Critique | Full document review with structured feedback |
| 6 — Finalization | Conclusions, citation formatting, consistency check |
Supported File & Evidence Types
| Input | What the Skill Does |
|---|---|
| PDF / DOCX thesis or paper | Reviews structure, consistency, and argumentation |
.csv / .xlsx data files | Checks sample size, distribution, and outliers |
.py / .R / .do / .m scripts | Reads analysis logic, verifies outputs match report |
| SPSS output / screenshots | Cross-references result tables with reported figures |
| SQL queries | Checks whether data collection matches description |
Installation
Prerequisites
- A Claude.ai account (free or paid)
- Node.js installed on your machine
Choose Your Installation Method
Option 1: Quick Install via npx (Recommended)
You can add this skill directly to your local project environment or global skill registry using the package runner:
npx skill add NguyenAn2812/scientific-research-helper
This command automatically pulls the instruction assets (SKILL.md and templates) from the remote repository and initializes them for your workspace.
Option 2: Manual Installation from Source
- Clone the repository
git clone https://github.com/NguyenAn2812/scientific-research-helper.git
cd scientific-research-helper
- Open Claude.ai
Navigate to Claude.ai and sign in to your account.
- Access the Skills panel
Click your avatar in the bottom-left corner, then select Skills (or Projects / Custom Instructions depending on your Claude interface) from the menu.
- Add the skill
Click Add Skill → Upload from file, then select SKILL.md from your local copy of the repository.
- Verify the installation
The skill should now appear in your Skills library. Start a new conversation and select Scientific Research Helper from the skill picker to begin.
Keeping the Skill Updated
To sync the latest improvements and template updates:
# If installed via manual clone
git pull origin main
# If managed via CLI tool
npx skill update NguyenAn2812/scientific-research-helper
Evaluation & Benchmark
To validate the efficiency of the Scientific Research Helper, we established a rigorous technical benchmark. For this evaluation, we selected DeepSeek-V4 (via open-source deployment) as our base language model.
Why DeepSeek-V4?
- Cost-Effective & Open-Source: It is currently one of the most powerful, highly accessible, and completely free/open-source models available, making it the perfect daily driver for students and independent researchers.
- Reasoning Capabilities: It provides an ideal baseline to test whether an external system prompt (Skill) can significantly amplify a model's architectural and mathematical verification capabilities.
Before vs. After Evaluation Matrix (IEEE Layout & PyTorch Code Verification)
The benchmark was conducted using a "trapped" IEEE paper draft paired with an actual Jupyter Notebook (.ipynb) containing silent bugs (e.g., unreported data dropping via .dropna(), model architecture mismatches like nn.GELU() vs. ReLU, and heavy validation overfitting logs).
The evaluation is based on a strict 1–5 scoring rubric measuring the model's accuracy, deep structural vision, and behavioral adherence:
| Evaluation Metrics (IEEE & Code Verification Tasks) | Environment | Score | Empirical Evidence from Model Output | Status |
|---|---|---|---|---|
| Task 1: Accuracy Mismatch DetectionPaper: claims 96.5% | Notebook: logs 78.5% | Before(Pure DeepSeek-V4) | 2/5 | Overlooked the final training log cell entirely. Simply parsed the plain text and praised the 96.5% paper claim. | ❌ Fail |
| After(DeepSeek + Skill) | 5/5 | Flagged under Clear contradictions: "Mâu thuẫn: Văn bản ghi đạt 96.5% nhưng dòng log số #3 trong Notebook chỉ đạt val_acc: 0.785". | Success | |
| Task 2: Dataset Sample Size ($N$) DropPaper: states N=10,000 | Notebook: drops to 8,450 | Before(Pure DeepSeek-V4) | 1/5 | Completely missed the data-reduction caused silently by the .dropna() method. | ❌ Fail |
| After(DeepSeek + Skill) | 4/5 | Successfully detected the sample size contraction. Prompted the user to update the methodology section (Section III) of the IEEE paper. | Success | |
| Task 3: Mathematical & Code AlignmentPaper: defines ReLU | Notebook: uses GELU | Before(Pure DeepSeek-V4) | 3/5 | Ignored the discrepancy, treating both activation functions as generic, interchangeable non-linear layers. | ⚠️ Marginal |
| After(DeepSeek + Skill) | 5/5 | Logged under Needs review: "Kiến trúc mạng trong PyTorch sử dụng nn.GELU() nhưng công thức toán ở phần III (IEEE Layout) lại định nghĩa ReLU". | Success | |
| Task 4: Structural Response AdherenceRequirement: Strict 3-tier Markdown layout | Before(Pure DeepSeek-V4) | 3/5 | Generated a long, essay-style narrative review. Required the researcher to manually dissect the feedback. | ⚠️ Text-heavy |
| After(DeepSeek + Skill) | 5/5 | Strictly adhered to the requested Markdown format: utilized structural rules (---), targeted bolding, and structured inventory tables. | Success |
Performance Improvement Formula
Using the quantified benchmark matrix above, we can measure the net performance gain:
$$\text{Improvement Rate} = \left( \frac{\sum \text{After Scores} - \sum \text{Before Scores}}{\sum \text{Before Scores}} \right) \times 100%$$
- Total Before Score: $2 + 1 + 3 + 3 = 9$
- Total After Score: $5 + 4 + 5 + 5 = 19$
- Net Skill Performance Boost: ~111.1% increase in structural code-to-paper verification accuracy.
Repository Structure
scientific-research-helper/
├── SKILL.md # Core skill instructions
├── scientific-research-helper.skill # skill kit
└── references/
├── research-proposal-template.md # Standard proposal structure
├── writing-review-checklist.md # Section-by-section review checklist
└── topic-idea-checklist.md # Topic feasibility evaluation checklist
Design Philosophy
- Guide, don't ghostwrite. The skill helps you think through your research — it won't write your thesis for you.
- Ask before assuming. When context is unclear, it asks one or two targeted questions rather than making wrong assumptions.
- Evidence first. For quantitative research, it always asks to verify numbers against actual outputs before giving feedback.
- No hallucinated references. If it can't find a source through search, it says so. It will never fabricate author names or paper titles.
Contributing
Contributions are welcome and appreciated! Whether you're fixing a typo, adding support for a new research methodology, or improving an existing workflow — your help makes this skill better for everyone.
Ways to Contribute
- Report a bug — Open an issue describing the problem, including the steps to reproduce it and your environment.
- Suggest a feature — Have an idea for a new workflow, document type, or discipline? Open an issue with the enhancement label.
- Submit a pull request — See something you can fix or improve? We'd love to review your PR.
Step-by-Step Contribution Guide
If you'd like to code or write templates for this repository, please follow these steps:
- Open an issue first — Describe what you'd like to change and why so the community can discuss it.
- Fork the repository — Create your own copy of the project on GitHub.
- Clone your fork locally
git clone https://github.com/NguyenAn2812/scientific-research-helper.git
cd scientific-research-helper
- Create a feature branch
git checkout -b feat/your-awesome-feature
- Make your changes & follow the Development Guidelines below.
- Commit and Push your changes
git add .
git commit -m "feat: add support for qualitative narrative analysis"
git push origin feat/your-awesome-feature
- Open a Pull Request — Navigate to the original repository and click "New Pull Request". Reference your initial issue in the PR description.
Development Guidelines
- Match the existing tone — The skill communicates with users in a structured, encouraging, and evidence-first style. New instructions should follow the same voice.
- Preserve stage awareness — If your contribution adds a new research stage or workflow, ensure it integrates cleanly with the existing stage system (Stages 1–6).
- Don't break existing workflows — Verify that your changes don't introduce regressions in the file review, writing plan, or literature search features.
- Document your additions — If you add new references or templates to the
references/directory, include a brief header explaining their purpose.
Code of Conduct
This project adheres to a simple standard: be respectful and constructive. Harassment, personal attacks, or unprofessional behavior will not be tolerated. All interactions — whether in issues, pull requests, or discussions — should foster a welcoming and collaborative environment.
License
MIT License — free to use, modify, and distribute.