Summary

This plugin addresses critical research-specific errors made by AI code assistants, which can lead to significant wasted compute and effort. Targeting researchers and power users, it provides guardrails and specialized skills to enhance AI-assisted workflows, ensuring accuracy in tasks like paper reproduction, experiment design, and debugging, thereby saving time and improving research integrity.

How It Works

The plugin operates as an extension for Claude Code, emphasizing a "methodology over scripts" approach. It equips the AI with research-specific skills, enabling it to generate tailored code based on the user's environment (e.g., wandb, local files). Core design principles include "human oversight first," integrating verification checkpoints, and delivering "actionable output" with ranked suggestions and specific fixes. Silent "Research Guardrails" proactively catch common AI blunders.

Quick Start & Requirements

• Install: claude plugin install phd-skills@phd-skills via the Claude plugin marketplace.
• Prerequisites: None explicitly stated beyond access to Claude Code. Optional notification setup requires ntfy, Slack, or email.
• Setup Time: Approximately 30 seconds for the optional /phd-skills:setup tour.
• Links: No direct documentation or demo links provided in the README.

Highlighted Details

• Paper Reproduction: Features a 7-stage skill to replicate ArXiv papers from URL to executable runs.
• Code-Paper Audit: Conducts a 5-dimensional parallel audit for paper-code consistency.
• Experiment Monitoring: Integrates with experiment tracking tools (wandb, neptune, etc.) and offers SSH notifications.
• Research Guardrails: Implements 11 hooks to prevent common AI errors, including unverified commands, fabricated paths, unreviewed figures, and research state loss.
• Skills Library: Offers commands like /xray, /factcheck, /gaps, /fortify, and auto-triggering skills for debugging, comparison, launching, and more.

Maintenance & Community

Developed by Fatih Cagatay Akyon, a researcher with extensive citations and patents. No other contributors, community channels (Discord/Slack), sponsorships, or partnerships are detailed in the provided text.

Licensing & Compatibility

• License: MIT.
• Compatibility: Permissive for research use, modification, and forking. No explicit restrictions noted for commercial applications.

Limitations & Caveats

The plugin is dependent on the Claude Code AI assistant. Its scope is focused on mitigating AI-induced errors within research workflows, rather than providing standalone research tools. While robust, its effectiveness relies on the underlying AI's capabilities and the user's specific research context.