Human-guided Research Agent Skill for Master and PhD students in CS, AI, math, and engineering: literature grounding, novelty gates, math formalization, experiment planning, reviewer simulation, and...
개요
Use agent research skills to read deeper, think wider, plan better, and review harder while the researcher stays in control. Academic Research Agent Skill helps Master, PhD, and independent researchers in technical disciplines collaborate with AI agents across the research lifecycle: scoping, literature review, research paper planning, novelty checking, mathematical formalization, experiment planning, reviewer simulation, and claim verification. It is designed for Claude-style skills, Claude Code workflows, ChatGPT/Codex-style agents, Gemini, and local LLM research assistants. It is not an autonomous paper factory or an "auto research paper" generator. The goal is not to remove the researcher. The goal is to give researchers a disciplined collaboration workflow so they can learn faster, explore more ideas, challenge weak assumptions earlier, and complete research with stronger evidence. - Master students learning how to structure a serious research project.
README
Academic Research Agent Skill
A Claude-compatible academic research agent skill for PhD and Master students: literature review, novelty checks, research Reality Gates, feasibility testing, experiment planning, reviewer simulation, and claim verification.
Use agent research skills to read deeper, think wider, plan better, and review harder while the researcher stays in control.
Academic Research Agent Skill helps Master, PhD, and independent researchers in technical disciplines collaborate with AI agents across the research lifecycle: scoping, literature review, research paper planning, novelty checking, mathematical formalization, experiment planning, reviewer simulation, and claim verification. It is designed for Claude-style skills, Claude Code workflows, ChatGPT/Codex-style agents, Gemini, and local LLM research assistants.
It is not an autonomous paper factory or an “auto research paper” generator. The goal is not to remove the researcher. The goal is to give researchers a disciplined collaboration workflow so they can learn faster, explore more ideas, challenge weak assumptions earlier, and complete research with stronger evidence.
Why Students Use It
| Need | How the skill helps |
|---|---|
| Start from a rough thesis idea | Turns broad topics into scoped research questions, non-goals, and contribution options. |
| Read literature more systematically | Forces source inspection, evidence extraction, and grounding against closest prior work. |
| Process real papers at scale | Supports paper download, PDF-to-Markdown conversion, figure/table extraction, and source analysis matrices. |
| Avoid weak novelty | Runs a novelty gate before implementation or drafting consumes too much time. |
| Catch polished but infeasible plans | Tests the real unit, metric, intervention, access path, valid yield, and joint claim dependencies before deep planning. |
| Learn research structure | Makes assumptions, gates, reviewer risks, and decisions explicit. |
| Prepare experiments | Builds risk plans, work breakdowns, pilot criteria, and execution artifacts. |
| Write stronger claims | Verifies claims against sources, formal artifacts, or experiment results. |
Who It Is For
- Master students learning how to structure a serious research project.
- PhD students managing literature, novelty, experiments, and writing pressure.
- Students in computer science, AI, machine learning, mathematics, engineering, and related technical fields.
- Research engineers turning rough ideas into executable experiments.
- Supervisors who want traceable AI-assisted research artifacts.
- Students and researchers who need a reusable research workflow.
What’s Included
| Area | Included |
|---|---|
| Skill entry | SKILL.md, CLAUDE.md, agents/openai.yaml |
| Commands | 16 prompt commands for scoping, ingestion, novelty, planning, review, and execution |
| References | Workflow, roles, source grounding, tool layer, novelty gate, experiments, language policy |
| Agent roles | Orchestrator, Strategist, Critic, Planner, Architect, Executor, DevOps |
| Examples | Topic brief and sample outputs for scope, literature grounding, novelty, and claims |
| Landing page | Static GitHub Pages site in docs/ |
| Visuals | Hero image, social preview, Mermaid diagrams, and image-generation prompts |
Core Promise
flowchart LR
H["Researcherjudgment, goals, approval"] A["Agent teamreading, drafting, critique, planning"]
A --> E["Evidence-traced artifacts"]
E --> H
H --> D["Better research decisions"]
The researcher provides direction, taste, constraints, and final judgment. The agents provide breadth, structure, critique, and execution support. Every important step creates an artifact that can be reviewed, revised, and learned from.
What This Skill Does
- Turns rough ideas into scoped research questions and contributions.
- Supports Claude skill for PhD student workflows, Claude skills for research paper planning, and agent research workflows.
- Forces source inspection before claims are used.
- Works with tool-assisted paper ingestion: download PDFs, convert them to Markdown, extract figures/tables, and build source analysis matrices.
- Grounds methods against closest prior work.
- Blocks shallow novelty with an explicit novelty gate.
- Requires the minimum mathematical definition needed to falsify a claim.
- Runs a Research Reality Gate before deep implementation planning.
- Separates technical smoke tests, scientific feasibility pilots, and full runs.
- Builds risk plans, work breakdowns, and code execution plans only for reality-cleared claims.
- Simulates reviewers before the paper is too expensive to fix.
- Verifies claims against sources or experiment artifacts.
- Supports configurable output language while keeping prompts stable in English.
Tool-Assisted Paper Ingestion
This skill is designed for real research workflows where papers are processed with tools, not only pasted into chat.
| Tool layer | What it enables |
|---|---|
| Paper downloader | Collect PDFs from source lists, arXiv links, DOI pages, or curated reading lists. |
| PDF-to-Markdown converter | Turn dense PDFs into readable Markdown notes that agents can inspect and cite cautiously. |
| Figure/table extractor | Read captions, crop figures/tables, and analyze visual evidence that may not appear in plain text. |
| Source analysis matrix | Compare papers by problem, method, dataset, metric, result, limitation, and relevance. |
| Claim tracer | Link draft claims back to inspected sources, figures, tables, or experiment artifacts. |
See references/tool_layer.md for the expected output contracts.
What This Skill Does Not Do
- It does not write a fake paper from thin air.
- It does not promise automatic research paper generation without human judgment, sources, methods, and verification.
- It does not invent citations, datasets, baselines, or results.
- It does not replace advisor feedback or human research judgment.
- It does not guarantee acceptance at any venue.
- It does not claim full autonomy.
Quick Start
Clone the repo into a research workspace used with Claude, ChatGPT, Gemini, local LLM agents, or another AI agent that can read repository instructions.
git clone
cd Academic-Research-Agent-Skill
cp config/language.example.yaml config/language.yaml
Then ask your assistant:
Use SKILL.md and CLAUDE.md. Help me develop this research idea:
"My research topic here"
Recommended first sequence:
/paper-scope
/pdf-ingest
/lit-ground
/math-formalize
/astar-novelty
/reality-gate
# If authorized: experimental-unit audit and feasibility pilot
/risk-plan
/code-exec-plan
/reviewer-sim
/claim-verify
Language Configuration
Prompts are written in English for consistency. Outputs can be localized.
output_language: "Vietnamese"
secondary_language: "English"
translation_mode: "technical-terms-in-english"
See docs/LANGUAGE_CUSTOMIZATION.md.
Workflow
flowchart TD
I["1. Ideahuman intent"] --> S["2. Scopequestion, contributions, non-goals"]
S --> G1{"Human gateIs this worth pursuing?"}
G1 -->|revise| S
G1 -->|approve| L["3. Source ingestionpapers, notes, code, datasets"]
L --> LG["4. Literature groundingclosest prior work and baselines"]
LG --> M["5. Math formalizationdefinitions, objective, assumptions"]
M --> N{"6. Novelty gatenot just method A on problem B"}
N -->|fail| S
N -->|conditional| F["Fix novelty or scope"]
F --> N
N -->|pass| G2{"7. Research Reality Gatewhat does evidence authorize?"}
G2 -->|block| E["Bounded evidence recovery"]
E --> G2
G2 -->|pilot only| X["8. Claim-eligible feasibility pilot"]
X --> G2
G2 -->|execution ready| P["9. Claim freeze, risk and work plan"]
P --> C["10. Approved code and experiment plan"]
C --> G3{"Full-run gatePilot and provenance sufficient?"}
G3 -->|no| X
G3 -->|yes| R["11. Approved run and reviewer simulation"]
R --> V["12. Claim verification"]
V --> O["13. Draft, revise, submit, or archive"]
Agent Collaboration Model
graph TB
H["Human Researcherdirection, constraints, approval"]
O["Orchestratorstate, routing, gates"]
S["Strategistscope, literature, math"]
C["Criticnovelty, reviews, claim checks"]
P["Plannerrisks, milestones, WBS"]
A["Architectcode and experiment design"]
E["Executorimplementation and pilots"]
D["DevOpsenvironment and remote runs"]
H O
O --> S
O --> C
O --> P
O --> A
O --> E
O --> D
S C
C --> H
P --> H
A --> H
E --> H
Repository Map
.
├── SKILL.md # Core skill entry point
├── CLAUDE.md # Agent session instructions
├── references/ # Skill references loaded when needed
├── .claude/commands/ # Slash-command style prompts
├── _agents/ # Role contracts, rules, workflows
├── config/ # Language configuration
├── docs/ # User-facing documentation
├── examples/ # Topic brief and sample output structure
└── assets/image-prompts/ # Prompts for generating repo visuals
Why This Is Different
| Compared with | Difference |
|---|---|
| Fully autonomous paper generators | Human approval is a core design feature, not a fallback. |
| General deep research tools | Focuses on academic contribution shaping, novelty, formalization, and review. |
| Prompt collections | Defines roles, gates, artifacts, and traceability rules. |
| Literature-only assistants | Connects literature to math, experiments, implementation, and claims. |
Expected Artifacts
02_Scope.md05_Lit_Grounding.md06_Math_Formalization.md20_Reality_Gate.md22_Experimental_Unit_Audit_Plan.md23_Feasibility_Pilot_Protocol.md10_Risk_Plan.md11_WorkBreakdown.md12_Code_Execution_Plan.md14_Agent_Brief_PhaseN.md15_Changelog.md19_Source_Analysis_Matrix.md- PDF-to-Markdown notes, figure/table reports, and download logs when tool-assisted ingestion is used
- Reviewer simulation and claim verification report
Visual Assets
The repository includes Mermaid diagrams and image-generation prompts:
- docs/visuals/human_agent_collaboration.mmd
- docs/visuals/research_lifecycle.mmd
- docs/visuals/artifact_pipeline.mmd
- assets/image-prompts/README.md
Documentation
- Getting Started
- Research Tools
- Positioning
- Use Cases
- Language Customization
- Visual Guide
- Image Generation Guide
- Competitive Analysis
- Launch Checklist
Recommended GitHub Description
Claude-compatible academic research agent skill for PhD/Master students: literature review, research paper planning, novelty checks, experiment planning, reviewer simulation, and claim verification.
Search Phrases This Repo Serves
This repository is relevant to searches such as Claude skill for PhD student, Claude skills for research paper, agent research skills, agent research workflow, academic research agent, AI research assistant for literature review, PhD research workflow, novelty check, claim verification, and auto research paper alternatives.
Safety Position
Academic Research Agent Skill is designed for evidence-traced collaboration. If a claim cannot be linked to an inspected source, formal artifact, or experiment result, it must be labeled as a hypothesis or removed.
Citation
If you use this repository in academic work, cite it as a human-guided academic research agent skill for Master and PhD students in technical fields.
추천 도구
다른 키워드를 입력하거나 필터를 제거해 보세요.
설치
npx skillfish add ngtiendong/academic-research-agent-skill