Claude 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
| 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 | bashChoose user-level (~/.claude/, recommended) or project-level (./.claude/) when prompted.
Selective — cherry-pick only the commands you want:
git clone https://github.com/Gaaaavin/claude-eureka.git /tmp/ce
cp /tmp/ce/commands/experiment.md ~/.claude/commands/
cp /tmp/ce/commands/debug.md ~/.claude/commands/
rm -rf /tmp/ceFrom a local clone — for contributors:
git clone https://github.com/Gaaaavin/claude-eureka.git
cd claude-eureka && ./install.shOptional, but unlock deeper capabilities:
# Weights & Biases — query experiments, runs, traces
claude mcp add wandb -- npx -y @anthropic-ai/mcp-wandb@latest
# GitHub — PRs, issues, code search
claude mcp add github -- npx -y @anthropic-ai/mcp-github@latestOpinionated defaults, fully escapable. Works great out of the box. Every skill is a markdown file you can edit or replace.
Zero runtime dependencies. No npm, no pip, no Docker. ML researchers have enough dependency hell. The entire project is .md files and one shell script.
Context is precious. Skills are concise by design. Claude is already smart — we give it your project's specific context, not a thousand-line system prompt.
Research-first, always. /debug knows about gradient explosions and CUDA OOM. /review catches research anti-patterns like data leakage and metric p-hacking. /scaffold generates PyTorch modules, not React components.
The fastest path to contributing:
- Use
/create-skillto author a command or skill locally - Test it in your own workflow for a few days
- Use
/contribute-skillto open a PR — it handles the packaging
See CONTRIBUTING.md for structure requirements, style guide, and CI checks.
Apache 2.0 — use freely, attribution appreciated.