paper-plan

The paper-plan skill automates the transition from raw research artifacts to a coherent, venue-ready paper outline. Designed for academic researchers and data scientists using agentic workflows, it processes multi-modal inputs including narrative documents (STORY.md), raw experiment logs, JSON figure data, and auto-review summaries. By synthesizing these inputs, the agent builds a comprehensive Claims-Evidence Matrix to ensure every section of your paper is grounded in empirical results.

• Automatically maps research claims to specific experiment evidence and metrics.

• Supports tailored structures for diverse venues like ICLR, NeurIPS, ICML, CVPR, and IEEE journals/conferences.

• Incorporates external writing principles and venue-specific checklists to ensure narrative flow and adherence to page limits.

• Generates section-by-section breakdown including abstract, introduction, method, and discussion, helping authors maintain a logical research story.

• Leverages cross-model collaboration, using an external reviewer model (via Codex MCP) to validate the outline's strength and logical consistency before writing begins.

• Prerequisites: Ensure research findings (findings.md), experiment logs, or AUTO_REVIEW.md are available in the project directory for accurate generation.

• Usage: Trigger with commands like '写大纲', 'paper outline', or 'plan the paper'.

• Customization: Users can override the target venue (e.g., --venue: NeurIPS) to automatically adjust MAX_PAGES and structure requirements.

• Context Management: The agent prefers compact files like IDEA_CANDIDATES.md to minimize context window consumption while maintaining high-quality generation.

• Integration: Designed to work seamlessly within the ARIS (Auto-claude-code-research-in-sleep) research framework, supporting various IDEs like Claude Code, Cursor, and Trae.