research-planning — Claude Code skill: research and paper plans

What research-planning does

The research-planning skill takes a topic, an idea description or a paper to reproduce and turns it into a plan. It starts by reading whatever papers, code or references you supply, pinning down the core research question and checking what resources are on hand: datasets, compute, existing code.

Planning runs in four stages. The overall plan covers methodology, key experiments and evaluation metrics. Architecture design maps file structure and system design with Mermaid class and sequence diagrams. Logic design breaks the work into tasks with dependencies, required packages and shared knowledge, and configuration collects hyperparameters plus a config.yaml. A separate pass lays out the paper section by section, from Abstract to Conclusion, with key points, figures and tables for each. The result is JSON holding the research question, methodology, task list, baselines, datasets, metrics and risks.

The skill reads two reference files from its own folder: references/planning-prompts.md and references/output-schemas.md. Vague input produces a vague plan, and by design the skill flags ambiguities instead of filling the gaps itself. Worth knowing before you run it: a fuzzy prompt means a rewrite and a second pass.