Agent Skills

Agent Skills are markdown files you can drop into your AI assistant. They turn the raw MCP tools into well-paved workflows — picking the right GPU for a workload, launching VMs safely with the right pre-flight checks, auditing spend, and cleaning up.

Skills are content-portable plain markdown — the same file works in Claude Code, Claude Desktop, Cursor, ChatGPT custom GPTs, and any other AI client.

Why use these

Skills give your AI assistant a head start. Instead of figuring out a workflow from scratch every time, the assistant follows a pre-written playbook with the right defaults baked in.

Concretely, with these skills installed your assistant will:

Recommend GPUs from our actual inventory and pricing — not generic advice from training data.
Run pre-flight checks before launching, so you don't see a 402 because billing wasn't configured.
Show you exactly what's about to be terminated — UUID, name, GPU, uptime, accumulated cost — and ask for confirmation before doing it.
Compute your hourly burn rate and recharge runway directly, instead of estimating.

Skills are optional. Everything still works without them — they just make the experience more reliable and consistent.

Available skills

`mc-pick-gpu` — GPU selection by workload

Helps the assistant pick the right GPU SKU from Massed Compute's inventory for what you're trying to do (LLM inference, fine-tuning, diffusion, training, HPC). Encodes the full GPU lineup, a capability ladder, a 25-row decision matrix, VRAM math, and known SKU caveats.

Download mc-pick-gpu.md

`mc-launch-vms` — Single and multi-VM launch playbook

Orchestrates the full launch flow: API key check, billing pre-flight, SSH-key check, GPU and image selection, single or multi-instance launches with naming conventions, coupon-aware launches, and post-launch SSH wiring.

Download mc-launch-vms.md

`mc-cost-control` — Spend audit and recharge runway

Per-instance burn rate, fleet hourly cost, long-running detection, and recharge runway estimation. Read-only — does not terminate.

Download mc-cost-control.md

`mc-safe-terminate` — Safe cleanup

Wraps termination with mandatory disclosure (UUID, name, GPU, uptime, accumulated cost) and explicit user confirmation. Supports batch-by-filter cleanup ("terminate all idle >24h") with full disclosure before any tool call.

Download mc-safe-terminate.md

How to install

These skills follow the Agent Skills open standard: each skill is a directory named after the skill, containing a SKILL.md file with the body. Claude Code, Codex, and claude.ai / Claude Desktop all use this format — they only differ in where the skills directory lives. Cursor and Windsurf don't have a native skills format; for those, you paste the body into their rules system instead.

One-time setup for each downloaded file

The downloads are flat .md files. Convert one to the standard layout once, then point any compatible client at it. For example, with mc-pick-gpu.md:

mkdir -p mc-pick-gpu
mv mc-pick-gpu.md mc-pick-gpu/SKILL.md

You now have a mc-pick-gpu/ directory with SKILL.md inside. Drop that directory into the right path for your client.

Claude Code

Scope	Path
Personal (all your projects)	`~/.claude/skills/mc-pick-gpu/SKILL.md`
Project (committed to repo)	`.claude/skills/mc-pick-gpu/SKILL.md`

Skills load on the next claude session start, or live within an existing session if you edit a file under a watched skills directory. Verify with /skills inside a session.

claude.ai / Claude Desktop

Skills upload through the UI, not a file path. Zip the skill directory first:

zip -r mc-pick-gpu.zip mc-pick-gpu/

Then in claude.ai or Claude Desktop, open Customize → Skills, click +, choose Upload a skill, and select the ZIP. claude.ai skills are scoped per workspace — there is no global option in the UI.

Codex

Codex CLI uses the same SKILL.md directory format. Both directories are scanned at session start:

Scope	Path
User (all your projects)	`~/.codex/skills/mc-pick-gpu/SKILL.md`
Project (committed to repo)	`.codex/skills/mc-pick-gpu/SKILL.md`

Cursor

Cursor uses rules, not skills. Paste the body of the downloaded file (the YAML frontmatter is harmless):

Scope	Where
Global	Cursor Settings → Rules → User Rules (paste the body in)
Project	`.cursor/rules/mc-pick-gpu.mdc` in the repo root

Windsurf

Same pattern as Cursor — Windsurf has rules, not skills:

Scope	Where
Global	Windsurf Settings → Rules → Global Rules (click Add Rule, paste the body in)
Workspace	`.windsurf/rules/mc-pick-gpu.md` in the repo root

ChatGPT custom GPTs / generic clients

Paste the body of the downloaded file into your GPT's "Instructions" field, or include it in the system prompt for any other MCP-capable assistant. The skill content is plain markdown with no client-specific syntax.