fcakyon/phd-skills

PhD Research Skills for Claude Code: paper reproduction, experiment design, paper review, result comparison and more.

What it is

phd-skills is a plugin for Claude Code (Anthropic's AI coding assistant) that adds research-specific guardrails and workflows for machine-learning researchers. It's a set of commands, auto-triggering skills, background agents, and safety scripts that make the AI assistant more reliable when doing scientific work like reproducing papers, debugging training runs, and comparing experiments.

The problem it solves

The author (an ML researcher) found that Claude Code makes costly, research-specific mistakes: launching training runs with wrong configs, analyzing the wrong dataset split, claiming tests passed without verifying, trusting generated numbers without looking at figures, and even running destructive commands on hallucinated paths. Each mistake can cost weeks of compute. This plugin adds guardrails to catch those mistakes before they happen.

How it works

Three layers, all installed via a single command with zero external dependencies:

  1. Commands (6 total) — explicit slash-commands like /phd-skills:reproduce arxiv 2508.12345 (reproduce a paper from arXiv) and /phd-skills:xray (audit a paper against its code and data across 5 dimensions).

  2. Skills (12 total) — auto-trigger when you describe a task in plain language. Say "why is my loss diverging?" and a Debug skill runs evidence-first probes; say "compare run alpha to baseline" and a Compare skill aligns experiments at the same epoch before judging.

  3. Guardrails (11 total) — silent background checks. Examples: a fresh-context "research peer" reviews conclusions against actual artifacts; a script blocks rm -rf on paths the AI may have hallucinated; another verifies that generated figures were actually looked at; a pre-flight checklist runs before long ML training launches.

There are also two background agents that Claude delegates to automatically: a paper-auditor that cross-checks paper claims against code/data in an isolated worktree, and an experiment-analyzer that reads results from wandb/neptune/tensorboard/mlflow.

Design philosophy

The README emphasizes methodology over scripts — skills teach the approach and Claude generates code for your specific setup (wandb, local files, etc.) rather than shipping rigid tools. It also stresses human oversight: every skill includes verification checkpoints because AI assistants tend to make premature claims.

Bottom line

A practical safety layer for anyone using Claude Code to do serious ML research — it doesn't add new AI capabilities, it makes the existing AI assistant less likely to waste your compute on avoidable mistakes.

Related

  • Project
    OpenImagingLab/AnyRecon

    AnyRecon is a 3D reconstruction framework that uses a video diffusion model to generate high-quality, arbitrary-view reconstructions of scenes from a set of capture views.

  • Project
    OliverDOU776/Few-step-probabilistic-glucose-forecasting-from-continuous-glucose-monitoring-and-meal-images

    A fast probabilistic forecasting tool that uses conditional rectified flow to predict glucose levels, transferring knowledge from large source datasets to small target datasets.

  • Project
    Softlandia-Ltd/vision-is-all-you-need

    A Vision RAG (V-RAG) demo that embeds PDF pages as images using a VLM to eliminate the need for text chunking during document retrieval.

  • Project
    Wakals/CoVT

    CoVT is a framework that enables Vision-Language Models to reason using continuous visual tokens, improving their spatial reasoning and geometric awareness by grounding semantic thoughts in perceptual cues.