Use Cursor for academic paper writing with citations
LLM Interfaces for coding are probably a generation ahead of what we have for prompting and interconnecting complex, disparate information. Task: The author conducted a crude experiment to construct an NSF Project Summary using Cursor. They set up a main directory with a markdown section file and a reference directory containing text versions of two papers. Using Agent mode with claude-3.7-sonnet, they applied two .cursorrules files: one for NSF requirements and one for academic writing standards. The author manually @'ed the writing rule file during final revision to ensure it was applied, acting as the primary editor and context manager.; tools: Cursor. Human role: The original post does not say how a person stayed involved.; limitations: Cursor lacks native PDF support, requiring external conversion to Markdown (e.g., via Datalab or Marker) which complicates context management. The tool is primarily optimized for coding, leading to issues with modifying or interrupting general text generation. Collaboration is difficult if peers use Google Docs or Word instead of Markdown/Git. AI output requires heavy human paraphrasing and critique to avoid hallucinations, and the workflow is only slightly faster than writing directly..
Tools
Published on Agent Work Atlas:
Overview
LLM Interfaces for coding are probably a generation ahead of what we have for prompting and interconnecting complex, disparate information.
Workflow
The author conducted a crude experiment to construct an NSF Project Summary using Cursor. They set up a main directory with a markdown section file and a reference directory containing text versions of two papers. Using Agent mode with claude-3.7-sonnet, they applied two .cursorrules files: one for NSF requirements and one for academic writing standards. The author manually @'ed the writing rule file during final revision to ensure it was applied, acting as the primary editor and context manager.
Reported result
The initial draft was described as 'not bad but not that good' compared to accepted NSF summaries. Switching to Claude thinking mode helped but introduced hallucinations and misinterpretations. The author found the process less painful than single-stream chatbots but noted that Cursor often interrupts generation or cannot be modified easily when outputting general content. No measured success metrics were reported, only qualitative feedback on draft quality and workflow friction.
Human role
The original post does not say how a person stayed involved.
Limitations
Cursor lacks native PDF support, requiring external conversion to Markdown (e.g., via Datalab or Marker) which complicates context management. The tool is primarily optimized for coding, leading to issues with modifying or interrupting general text generation. Collaboration is difficult if peers use Google Docs or Word instead of Markdown/Git. AI output requires heavy human paraphrasing and critique to avoid hallucinations, and the workflow is only slightly faster than writing directly.
Try it
Act as an academic writing assistant using Cursor. I will provide a grant proposal template (Markdown) and two source papers (converted to text/Markdown). 1. Create two .cursorrules files: one defining the specific grant requirements (e.g., NSF Project Summary structure) and one defining academic writing style guidelines. 2. Place the template in the main directory and sources in a 'references' folder. 3. Use Agent mode with a high-reasoning model (e.g., Claude 3.7 Sonnet with thinking enabled). 4. Prompt the agent to draft the summary using the rules and references. Explicitly @mention the writing rules file if the initial draft ignores them. 5. Review the output for hallucinations. If the draft is weak, ask the agent to critique its own output against the rules. 6. Output the final draft in Markdown. Note: Do not send any real sensitive data. Use dummy papers for this dry run. Confirm the draft structure matches the template before proceeding to full content generation.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.