phd-skills

A community Claude Code plugin by Fatih Cagatay Akyon (fcakyon) that catches the research mistakes AI agents make before they cost weeks of compute. It bundles twelve skills, research commands, two helper agents and a set of guardrail hooks, each traced to a real mistake from his PhD work, such as restarting a long training job without diffing its config or trusting a figure nobody looked at. What you can do with it - Reproduce a paper from its arXiv link through to replication runs - Debug a failing, diverging or out-of-memory run with evidence-first probes - Compare a run to its baseline at the same epoch, across wandb, Neptune, TensorBoard, MLflow or local logs - Launch a training run with a pre-flight checklist - Design ablations, find related papers and analyze dataset bias - Check that a paper's numbers match the code, verify BibTeX entries and citations, set up LaTeX for a venue, and prepare for reviewer questions How it works - Skills load on their own when you describe the task, such as "why is my loss diverging?" or "compare run alpha to baseline". - Commands include /phd-skills:xray (audit a paper against its code and data), /phd-skills:factcheck, /phd-skills:gaps and /phd-skills:fortify. - Guardrail hooks run in the background: they flag destructive commands on made-up paths, in-place edits over SSH, unreviewed figures, LaTeX errors and missing citation checks, and save research state before the context fills up. - Install it from the maker's own marketplace with claude plugin marketplace add fcakyon/phd-skills, then claude plugin install phd-skills@phd-skills. Good to know - No MCP server needed. Paper search uses the maker's Paper Search CLI (uv and Python 3.10 or newer); other skills don't need it. - /phd-skills:setup is optional and turns on extras such as ntfy, Slack or email notifications and LaTeX. - MIT licensed.