| name | run_mode |
| description | Use this skill for experiment run drafting when the user starts with `Run Python:`, `Slurm job:`, or `Create run:`. It applies when Codex should prepare run-specific code, scaffolding, and execution instructions without launching long-running jobs by default. |
| metadata | {"short-description":"Draft experiment runs safely"} |
Run Mode
Use this skill when the user is asking to create or prepare an experiment run.
Triggers
Apply this skill when the request starts with:
Run Python:
Slurm job:
Create run:
Objective
Prepare run-specific code, scaffolding, and execution instructions without launching long jobs by default.
Run ID Convention
Every run must use:
MM-DD-YY__<topic>__<variant>-vN
Generate the run_id automatically if the user does not provide one.
Standard Workflow
- Clarify the objective in 1-2 sentences and list assumptions.
- Propose minimal code changes covering core code and any helper files.
- Apply the required code changes in the repo.
- Create or update the run scaffold under
runs/<run_id>/.
The scaffold should contain:
prompt.md copied from the user request
src/ for run-specific helpers or wrappers
out/ as the runtime output root
summary.md for the run summary written alongside the run outputs. This will be created when summarize_run is called for this run. Don't create an empty summary.md file in advance.
Standard output paths:
- Logs:
runs/<run_id>/out/logs/
- Artifacts:
runs/<run_id>/out/artifacts/
Do not add .gitkeep files for run folders.
If the repo also contains weekly documentation, keep those reports under ./weekly/.
Execution Rules
- You may run only safe, fast checks such as linting, imports, or a small unit test.
- Do not run long training or evaluation jobs automatically.
- Do not submit jobs with
sbatch automatically.
Run Python Mode
When the user starts with Run Python::
- Draft and apply the code changes.
- Stop and ask the user to confirm execution.
- After confirmation, run the Python command and echo the exact command used.
Slurm Job Mode
When the user starts with Slurm job::
- Draft and apply the code changes.
- Create an
sbatch script in the run folder. If the repo already has a Slurm template, follow that template. Otherwise write a minimal script directly.
- Save the script to
runs/<run_id>/<run_id>.sbatch.
- Fill in placeholders or script fields for the job and run specifics.
- Do not submit the job.
- Provide the exact
sbatch runs/<run_id>/<run_id>.sbatch command.