Thesis Specialist: An Academic Paper Writing Assistant Based on the Folder-as-Agent Architecture

This article introduces the Thesis Specialist project, an innovative platform designed specifically for academic paper writing. It corely adopts the Folder-as-Agent architecture, organizing AI agent workflows through folder structures, supporting multiple large language models, and providing structured writing assistance for researchers. It addresses the context management challenges of traditional AI interaction modes in complex, long-cycle writing tasks.

Academic paper writing involves multiple links such as literature review, method design, and data analysis, which is a complex intellectual activity. The maturity of LLM technology has made AI-assisted writing possible, but traditional API calls or chat interface modes face challenges in dialogue history management and context maintenance in long-cycle projects, so a more effective collaboration architecture is urgently needed.

The Folder-as-Agent architecture treats folders as executable AI agents, where each folder contains all the context required for the task (instructions, references, templates, etc.). Its advantages include: state persistence (progress is saved as files without loss), composability (nested links to build complex workflows), portability (cross-model/platform migration), and auditability (recording intermediate products and modification history).

Thesis Specialist designs modules for academic writing scenarios: literature management (PDF analysis to generate abstract maps), research question refinement (organizing argumentation logic), methodology design (recommending methods and pointing out defects), data analysis and visualization (code assistance and result interpretation), writing polishing (adjusting academic norms), and citation format management (automatically organizing APA/MLA and other formats).

The project supports mainstream models such as Doubao, Claude, GPT, and Gemini. It shields underlying differences through standardized folder interfaces, allowing users to switch seamlessly. Characteristics of different models: GPT is balanced and general-purpose, Claude has long-context security, Gemini is multi-modal, and Doubao is Chinese-friendly; this design reduces vendor lock-in risks.

AI assistance needs to distinguish between assistance and replacement: acceptable uses (polishing, idea organization, code assistance, etc.); uses requiring caution (full-paragraph generation, conclusion derivation, innovative ideas). Academic institutions have formulated AI usage policies, and the platform guides responsible use by labeling AI-assisted parts and providing integrity checklists.

Limitations: insufficient domain expertise (lack of depth in sub-fields), factual accuracy issues (AI hallucinations), and innovation limitations (difficulty in breaking through). Optimization directions: integrating academic databases to ensure accuracy, fine-tuning models with domain experts, formulating human-machine collaboration guidelines, and developing hallucination detection tools.

Thesis Specialist is a beneficial attempt to transform AI applications from "toys" to "tools". It is recommended that researchers position AI as augmented intelligence, using it to handle tedious tasks and devoting cognitive resources to creative activities to promote academic progress rather than threatening integrity.