AgentResearchSkill

AgentResearchSkill is a research-oriented cognitive workflow architecture for structured, LLM-assisted academic work.

This repository is not a prompt collection and not a plug-and-play agent. It focuses on externalizing thinking, decision-making, and research posture when working with large language models.

---

Why This Exists

Most LLM-based research workflows fail in one of two ways:

1. Outputs are fluent but shallow — mostly summarization.
2. Outputs are structured but rigid — creativity is suppressed.

AgentResearchSkill is designed to address this gap by introducing a Thinking Layer and Contribution Layer that sit above tools and workflows.

The goal is not to make LLMs smarter, but to make human–LLM collaboration more reliable, inspectable, and academic-grade.

---

Core Ideas

1. Tools do not think

Each tool has a single responsibility:

• search
• categorize
• extract
• generate

Tools never judge correctness or relevance.

---

2. Thinking is externalized

Instead of relying on hidden chain-of-thought, this system externalizes reasoning as:

• thinking modes
• decision matrices
• quality gates
• research profiles

All decisions are observable and debuggable.

---

3. Research posture is a first-class concept

The system supports different research intentions:

• review

  • structured summarization
  • taxonomy and trend analysis

• academic

  • conceptual abstraction
  • formalization
  • hypothesis generation

Research depth is a strategy choice, not an accident.

---

4. Contribution is separated from reporting

A dedicated Contribution Generation step is introduced to:

• propose new concepts
• express structural relationships
• generate testable hypotheses

This step does not introduce new data and does not claim scientific truth.

It enables academic-style reasoning without hallucination.

---

5. Technical Depth & Fidelity

Depth is not an option; it is a requirement. The system enforces:

• Mathematical Rigor: Every core finding must include its technical mechanism.
• Content Parity: Strictly forbidding information loss between draft and final output.
• Exemplar-Driven Generation: Using established examples (see Examples/) as mandatory benchmarks for quality.

---

System Structure

• Workflow Layer → execution order
• Tool Layer → concrete operations
• Thinking Layer → reasoning strategy & routing
• Contribution Layer → academic abstraction
• Quality Gates → stability & integrity
• Delivery Layer → final output formatting & export

Each layer has a single responsibility and can evolve independently.

This system guarantees a complete end-to-end workflow, from information gathering to final deliverable output.

---

Example Use Case

Topic: Vision-based defect detection
Time range: last 5 years
Profile: academic
Contribution intensity: 1
Final Output: academic research report

Result:

• structured literature table
• conceptual framework
• formal abstraction of method trade-offs
• research-style report draft

---

Examples

• Academic Research Report: Generative Data Quality Assessment (2016-2026)

---

Academic Mode Output Sample

When the system is set to Academic Mode, it goes beyond simple summarization to provide deep technical analysis, formal definitions, and hypothetical research contributions.

Key Characteristics:

• Formalization: Expressing relationships through mathematical notations and theorems.
• Conceptual Abstraction: Proposing new frameworks (e.g., "Synergy-Risk Triangle").
• Structural Depth: Detailed methodology and taxonomy systems.

Snippet from Example:

4. Formal Academic Contributions

4.1 The Synergy-Risk Triangle We propose a non-linear dynamical relationship model. Let F be Fidelity, U be Utility, and P be Privacy protection:

• Theorem: There exists a parameter set $ \theta $ such that after a specific threshold, $\Delta F \propto \Delta U \propto -\Delta P $.
• Conclusion: A high-quality evaluation framework should not pursue the maximization of a single indicator but seek a triangular balance.

4.2 Judge-Consistency Metric (LC) For emerging technologies like G-Eval based on Chain-of-Thought (CoT), this report formally defines the LC metric: $$ LC = 1 - \frac{\text{Var}(S_{LLM})}{\text{Var}(S_{Human})} $$

---

Intended Audience

This project is intended for:

• researchers
• research engineers
• advanced LLM users
• system and framework designers

It is not intended as a beginner tutorial or turnkey agent system.

---

What This Project Is Not

• not an AutoGPT-style autonomous agent
• not a benchmark or evaluation framework
• not a claim of scientific novelty

It is a design pattern for trustworthy LLM-assisted research.

---

Status

• Workflow Layer: stable
• Tool Specs: stable
• Thinking Layer: v1
• Contribution Layer: v1
• Quality Gates: v1.1 (Depth-Aware)
• Technical Depth Policy: v1.1

This system is designed to evolve.

---

License

Open for research and experimentation.