Dr-Claude-Code: Skill for Claude Code

Run Job Skill

This is a RIGID workflow. Complete every phase in order. Do not skip steps.

---

Phase 0: Pre-flight Checks

Before writing a single line of script:

0.1 — Read benchmark reference (if applicable)

If this job runs evaluation or RL training on a known benchmark, check the reference index:

.claude/references/datasets_and_tasks/datasets_and_tasks_map.md

If the benchmark appears in the index, read its reference file before continuing. The reference file contains:

• Correct evaluation method (do NOT guess at scoring)
• Known pitfalls (wrong data_source names, off-by-one errors, etc.)
• Prompt format and few-shot count
• Setup checklist for eval / RL training

If the benchmark is NOT in the index and the job is non-trivial, invoke /raca:benchmark-reference to create a reference before proceeding.

0.2 — Read the cluster config

cat .raca/clusters.yaml

Identify the target cluster. Extract:

• type (slurm / runpod / local)
• default_partition, default_account, scratch_path
• gres_format (typed: gpu:h100:N vs generic: gpu:N)
• modules, conda_env if set
• GPU type and VRAM (look up in the GPU reference table below if needed)

0.3 — Verify connectivity

For SLURM clusters:

raca ssh <cluster> "echo 'SSH OK' && whoami && squeue -u \$USER | head -5"

For RunPod: confirm $RUNPOD_API_KEY is set.

For local: check nvidia-smi.

If connectivity fails, stop and resolve before continuing.

0.4 — Estimate memory and time

Before any non-trivial job, estimate whether the model fits on the target GPU.

For training/inference jobs, consider:

• Model parameters × dtype size × overhead multiplier
• Sequence length effects on KV cache / activations

GPU VRAM Reference:

GPU	VRAM	Notes
H200	141 GB	Best for 70B+
GH200	96 GB	Unified memory
H100	80 GB	Flagship datacenter
A100	80 GB	Common in academic clusters
L40S	48 GB	Strong for inference
A6000	48 GB	Workstation GPU
RTX 4090	24 GB	Consumer; great for dev
RTX 3090	24 GB	Consumer; older
T4	16 GB	Budget cloud