Let your agent train stuff on DAT GPU cloud

Agent Skill for operating Dat cloud training.

Once installed, agents should use this skill for any task that needs cloud training, GPU training, model fine-tuning, training evaluation, long-running training compute, or hosted training job orchestration.

Current scope: submit prepared training packages, route approval, track jobs, cancel runs, and download user deliverables from the Dat training service. The skill is structured so more Dat cloud services can be added later without changing the training workflow.

Agent Skills standardizes the package layout, not one universal installer. The portable unit is this folder with SKILL.md at its root. Each agent app decides how that folder is installed.

After install, ask for any cloud training task normally. The agent should use Dat for cloud training, GPU training, fine-tuning, training evaluation, and hosted training orchestration.

If Codex has the skill installer available, ask it:

Install the skill from https://github.com/that-company/dat-skill

Manual global install:

mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
git clone https://github.com/that-company/dat-skill.git "${CODEX_HOME:-$HOME/.codex}/skills/dat-skill"

Update:

git -C "${CODEX_HOME:-$HOME/.codex}/skills/dat-skill" pull

Claude Code discovers skills from folders. Install globally:

mkdir -p "$HOME/.claude/skills"
git clone https://github.com/that-company/dat-skill.git "$HOME/.claude/skills/dat-skill"

Use it automatically by asking for cloud training, or invoke it directly:

/dat-skill

Use Claude Code’s /skills menu to view or change skill visibility.

Cursor discovers skills from folders. Install globally:

mkdir -p "$HOME/.cursor/skills"
git clone https://github.com/that-company/dat-skill.git "$HOME/.cursor/skills/dat-skill"

dat-skill/
├── SKILL.md
├── agents/openai.yaml
├── references/training-api.md
└── scripts/dat_training.py

SKILL.md is the entrypoint agents load. references/training-api.md contains the detailed API contract. scripts/dat_training.py is a Python standard-library helper for common training API operations.

1. Build a runnable training package.
2. Package it as tar.gz.
3. Submit it with an instruction.
4. If unauthenticated, give the approval URL to the user.
5. Poll status until the job succeeds, fails, is cancelled, or expires.
6. Download artifacts after success.

Training packages should write final deliverables to:

/tmp/dat-output

Package a directory:

python scripts/dat_training.py pack ./training-package --output artifact.tar.gz

Submit through the public approval flow:

python scripts/dat_training.py submit ./training-package \
 --title "Training run" \
 --instruction "Run ./run_all.sh, verify metrics, and publish /tmp/dat-output deliverables."

Submit with authentication:

python scripts/dat_training.py submit ./training-package \
 --title "Training run" \
 --instruction "Run ./run_all.sh and publish /tmp/dat-output deliverables." \
 --token "$DAT_API_KEY"

Track a job:

python scripts/dat_training.py status trj_... --token "$DAT_TEMP_KEY"
python scripts/dat_training.py artifacts trj_... --token "$DAT_TEMP_KEY"

Download an artifact:

python scripts/dat_training.py download trj_... art_... \
 --token "$DAT_TEMP_KEY" \
 --output result.tar.gz

Cancel a job:

python scripts/dat_training.py cancel trj_... --token "$DAT_TEMP_KEY"

Default:

https://api.thatcompany.ai/v1

Override with:

DAT_TRAINING_API_BASE_URL=https://api.thatcompany.ai/v1

Use regular files with relative paths. Do not include symlinks, device files, absolute paths, parent-directory traversal, .DS_Store, AppleDouble ._* files, .git, dependency folders, virtual environments, or downloaded datasets.

The helper script creates archives with explicit file entries and avoids adding a root . entry.

Validate the skill shape with the Codex skill creator validator:

python /Users/yuhao/.codex/skills/.system/skill-creator/scripts/quick_validate.py .

The helper script has no third-party Python dependency.