Custom Skills

The built-in SKILL.md covers general GPU provisioning. You can create custom skills for specific workflows by following the same Agent Skills standard.

Skill structure

Each custom skill is a directory with a SKILL.md file:

.agents/skills/
├── swm-gpu-workflow/     # built-in swm skill
│   └── SKILL.md
└── deploy-model/         # your custom skill
    └── SKILL.md

Use YAML frontmatter so agents know when to activate it:

---
name: deploy-model
description: >
  Use this skill when the user wants to deploy a HuggingFace model
  to a vLLM inference server on a cloud GPU.
---

Example: Auto-deploy fine-tuned model

---
name: deploy-model
description: >
  Use this skill when the user wants to deploy a HuggingFace model to vLLM
  on a cloud GPU. Do NOT use for training or fine-tuning tasks.
---

# Deploy Model to vLLM

## Steps

1. Search for GPU with enough VRAM: `swm gpus -g <class> --sort price`
2. Create pod: `swm pod create -p <provider> -g <gpu> -n <model-name> -y`
3. Install vLLM: `swm setup install vllm <pod-id>`
4. Pull model: `swm models pull <pod-id> <model-path>`
5. Set as active: `swm models set <pod-id> <model-path> --restart`
6. Verify: `swm run <pod-id> "curl -s localhost:8000/v1/models"`
7. Enable guard: `swm guard enable <pod-id> --policy auto-down --idle 30m`
8. Report URL and pod ID to user

Example: Batch training pipeline

---
name: batch-train
description: >
  Use this skill when the user wants to fine-tune a model with Axolotl.
  Provisions a GPU, trains, pushes checkpoints, and auto-terminates.
---

# Batch Training Pipeline

## Steps

1. Create pod with auto-down: `swm pod create -p <provider> -g <gpu> -n train --lifecycle auto-down -y`
2. Pull workspace: `swm sync pull <pod-id>`
3. Install Axolotl: `swm setup install axolotl <pod-id>`
4. Start training: `swm run <pod-id> "cd /workspace && axolotl train config.yml"`
5. Push checkpoints: `swm sync push <pod-id>`
6. Pod auto-terminates when training completes (GPU drops to 0%)

## Verify

- Check training logs: `swm run <pod-id> "tail -50 /workspace/axolotl/logs/latest.log"`
- Check GPU utilization: `swm run <pod-id> "nvidia-smi"`

Best practices

Follow the same phase structure as the built-in skill:

1. State check — Look for existing resources (swm pod list)
2. Clarify — Ask user for requirements
3. Execute — Run swm commands with -y flags
4. Verify — Check that everything works before handing off
5. Hand off — Report results and next steps

Keep your SKILL.md under ~5,000 tokens. If you need more detail, use a references/ directory:

deploy-model/
├── SKILL.md
└── references/
    └── vram-requirements.md

The agent loads references on demand, keeping context lean.