Claude Code skills for ML/AI researchers
npx claudepluginhub gaaaavin/claude-eurekaClaude Code skills for ML/AI researchers. Commands for experiments, debugging, visualization, and SLURM jobs.
GaaaavinClaude Code skills for ML/AI researchers.
Stop re-teaching Claude your project every session. Eureka auto-detects your stack, manages experiments, debugs training, and learns as you work.
/plugin marketplace add Gaaaavin/claude-eureka
/plugin install claude-eureka@claude-eureka
Then inside Claude Code: /init-eureka
Claude Code is Anthropic's AI coding agent. Skills extend it with slash commands and passive triggers. claude-eureka is a curated skill pack built specifically for ML/AI research workflows.
Out of the box you get commands for experiments, debugging, code review, SLURM job submission, and publication-quality plots — all assuming you know PyTorch and care about research velocity, not boilerplate.
Before eureka: After /init-eureka:
"Here's my project Claude already knows:
structure..." ✓ PyTorch + Lightning + Hydra
"I use Hydra for ✓ experiment layout in runs/
config..." ✓ SLURM cluster + GPU types
"My runs go in..." ✓ active branches and open TODOs
"Oh, and I'm ✓ NaN bug you fixed last week
working on..."
1. Install (30 seconds)
Inside Claude Code:
/plugin marketplace add Gaaaavin/claude-eureka
/plugin install claude-eureka@claude-eureka
This registers the eureka marketplace and installs the plugin. No npm, no pip — it's markdown files and one shell script.
2. Initialize your project
Open Claude Code in your project directory and run:
/init-eureka
This scans your repo, detects your stack (PyTorch, Lightning, Hydra, W&B, SLURM, etc.), and writes a tailored CLAUDE.md. Claude reads it on every prompt. Your context is set.
3. Get to work
/experiment baseline --lr 1e-4 --batch 64
/debug ← training loss exploded
/viz runs/ ← generate paper-quality figures
/submit-job train.py ← SLURM submission
/slash invocations)| Command | What it does |
|---|---|
/init-eureka | Scan project → generate tailored CLAUDE.md |
/refresh-context | Re-detect stack, update auto-sections, keep your edits |
/experiment | Create, launch, track, and log experiments |
/debug | Root-cause debugging — investigate first, patch second |
/review | Code review with YAGNI/KISS + ML anti-pattern detection |
/scaffold | Boilerplate: model, dataset, trainer, config, SLURM script |
/viz | Publication-quality figures from experiment outputs |
/notebook | Structured Jupyter analysis notebooks |
/submit-job | SLURM submission, status monitoring, log tailing |
/create-skill | Author new skills or commands (guided) |
/contribute-skill | Package a skill and open a PR to this repo |
/update-eureka | Pull latest commands and skills from GitHub |
| Skill | Activates on |
|---|---|
research-debugging | Errors, exceptions, NaN, OOM, CUDA errors, tracebacks |
auto-memory | "remember", conventions, experiment completions, bug resolutions |
Eureka uses a tiered context architecture so Claude gets exactly what it needs without wasting tokens:
CLAUDE.md (~50 lines, loaded every prompt)
│ Project identity, stack, key paths, active work state.
│
└── .claude/context/*.md (loaded on demand)
Experiments, architecture decisions, resolved bugs,
team conventions. Claude loads relevant files per query.
|
└── auto-memory (agent-maintained)
Results and learnings written back automatically.
Your second session is smarter than your first.
/init-eureka populates tiers 1 and 2. The auto-memory skill fills tier 3 over time. /refresh-context re-runs detection to keep auto-generated sections current while preserving your manual edits.
Plugin install — recommended, native Claude Code integration:
/plugin marketplace add Gaaaavin/claude-eureka
/plugin install claude-eureka@claude-eureka
Update anytime with:
/plugin update claude-eureka@claude-eureka
Alternatively (older Claude Code versions):
curl -fsSL https://raw.githubusercontent.com/Gaaaavin/claude-eureka/main/install.sh | bash
Choose user-level (~/.claude/, recommended) or project-level (./.claude/) when prompted.