Setup Skill for Claude

# Last few long-job launches the plugin would have flagged
recent_launches=$(history 200 2> /dev/null | grep -E '(python.*train|accelerate launch|torchrun|deepspeed|sbatch|tmux.*python)' | tail -5)

# What the agentic Stop hook would orient on for this project
project_orientation=""
[ -f README.md ] && project_orientation+="README.md found  "
[ -f CLAUDE.md ] && project_orientation+="CLAUDE.md found  "
[ -f .claude/CLAUDE.md ] && project_orientation+=".claude/CLAUDE.md found  "
[ -f AGENTS.md ] && project_orientation+="AGENTS.md found  "

Lab Rat auto-detected the following:

Recent launches:    <count> training-job patterns in shell history
Project orientation: ${project_orientation:-no README / CLAUDE.md / AGENTS.md found at project root}

This is what the hooks and agentic Stop hook will use. No configuration needed.
The plugin works on the first prompt. Hooks fire silently when there is nothing to flag.

Want to configure any extras? (notifications, allowlist, CLAUDE.md, LaTeX) [y/N]

Which extras would you like to set up?

1. Notifications: get alerts when long tasks complete (ntfy / slack / email)
2. CLI Allowlist: auto-approve safe commands used in research workflows
3. Research CLAUDE.md: add research integrity rules to your project
4. LaTeX Environment: detect and configure LaTeX / BibTeX toolchain

Enter numbers separated by commas (e.g., 1,3,4), or "all":

How would you like to receive notifications?

1. ntfy.sh: free, works over SSH, no account needed
2. Slack: webhook to a channel
3. Email: via sendmail / msmtp

Enter number:

{
  "allowedTools": [
    "Read",
    "Glob",
    "Grep",
    "Bash(pdflatex *)",
    "Bash(biber *)",
    "Bash(bibtex *)",
    "Bash(python -c *)",
    "Bash(uv run *)",
    "Bash(git status)",
    "Bash(git diff *)",
    "Bash(git log *)",
    "Bash(nvidia-smi)",
    "Bash(htop)",
    "Bash(df -h)",
    "Bash(wc -l *)",
    "Bash(du -sh *)"
  ]
}

## Research Integrity Rules

- Code is the source of truth. When paper text and code disagree, the code is correct
  unless the user explicitly states otherwise.
- Never state a numerical result without tracing it to a specific file, log, or code output.
  If you cannot find the source, say so explicitly.
- Never assume domain-specific technical behavior. Verify against code before making claims
  about how methods, losses, architectures, or training procedures work.
- When editing paper text, change ONLY what was requested. Do not remove existing content,
  add unrequested sections, or "improve" surrounding text unless asked.
- Citation integrity: verify author names, venue, and year against DBLP before adding or
  modifying any BibTeX entry. Flag any citation detail you cannot verify.
- Edit code locally, commit, push, then `git pull` on the GPU host. Never edit tracked code
  in place over SSH.