Initialize Research Project

Set up a new research project in the current working directory.

• Project name: basename $(pwd) (e.g., my-cool-project)
• Package name: hyphens replaced by underscores (e.g., my_cool_project)
• Description: $ARGUMENTS if provided, else "Research project"
• Assets: ~/.claude/skills/init-project/assets/
• Cluster info: ~/.claude/skills/slurm-info-summary/references/slurm-cluster-summary.md

Steps

1. Derive names

PROJECT_NAME=$(basename $(pwd))
PACKAGE_NAME=$(echo "$PROJECT_NAME" | tr '-' '_')
ASSETS="$HOME/.claude/skills/init-project/assets"

2. Create directory structure

mkdir -p notebooks "src/${PACKAGE_NAME}" attachements research_notes .claude/skills/report
touch "src/${PACKAGE_NAME}/__init__.py"

3. Copy static files (no substitution needed)

cp "${ASSETS}/report-SKILL.md" .claude/skills/report/SKILL.md
cp "${ASSETS}/plotting.py" "src/${PACKAGE_NAME}/plotting.py"

4. Copy and substitute CLAUDE.md

cp "${ASSETS}/CLAUDE.md.template" CLAUDE.md
sed -i "s|{{PROJECT_NAME}}|${PROJECT_NAME}|g; s|{{PACKAGE_NAME}}|${PACKAGE_NAME}|g" CLAUDE.md

5. Copy and substitute pyproject.toml

Use the skill argument ($ARGUMENTS) as description. If empty, use "Research project".

DESCRIPTION="<$ARGUMENTS or 'Research project'>"
cp "${ASSETS}/pyproject.toml.template" pyproject.toml
sed -i "s|{{PROJECT_NAME}}|${PROJECT_NAME}|g; s|{{PACKAGE_NAME}}|${PACKAGE_NAME}|g; s|{{DESCRIPTION}}|${DESCRIPTION}|g" pyproject.toml

6. Copy and substitute .mcp.json from template + cluster info

1. Read cluster summary at ~/.claude/skills/slurm-info-summary/references/slurm-cluster-summary.md.

   • If it does not exist, tell the user: “Run /slurm-info-summary first to gather cluster info.” and stop.

2. Extract values for a single-GPU shared notebook job from the summary:

   Variable	What to look for	Example (Raven)
   PARTITION	Shared GPU partition (1 node, shared access, has GPUs)	gpu1
   GPU_TYPE	GPU model, lowercase	a100
   CPUS	Physical cores per GPU (total physical cores / GPUs per node)	18
   MEM_MB	Memory per GPU in MB (convert from GB if needed)	125000
   TIME	Max walltime for that partition	1-00:00:00

3. Detect CUDA module:

   CUDA_MODULE=$(module avail cuda 2>&1 | grep -oP 'cuda/[\d.]+' | sort -V | tail -1)
   

   • If found: use the module name (e.g., cuda/12.6)
   • If empty: use empty string ""

4. Copy template and substitute:

   cp "${ASSETS}/.mcp.json.template" .mcp.json
   sed -i \
     -e "s|{{PARTITION}}|${PARTITION}|g" \
     -e "s|{{GPU_GRES}}|gpu:${GPU_TYPE}:1|g" \
     -e "s|{{CPUS_PER_TASK}}|${CPUS}|g" \
     -e "s|{{MEM_MB}}|${MEM_MB}|g" \
     -e "s|{{TIME}}|${TIME}|g" \
     -e "s|{{CUDA_MODULE}}|${CUDA_MODULE}|g" \
     .mcp.json
   

7. Print summary

Show what was created and next steps:

• uv tool install git+https://github.com/kdkyum/jlab-mcp.git (one-time, if not already installed)
• uv sync to install project dependencies
• jlab-mcp start in a separate terminal to launch JupyterLab on a GPU node
• Start Claude Code in the project directory