Unsloth Workflows

Use this skill as a framework guide and a shell workflow runner. Keep three modes distinct:

• Discuss: explain concepts or inspect a request. Do not install, write files, download models, or train.
• Prepare: create reviewable configuration or task-specific scripts. Do not install dependencies or start GPU work.
• Execute: perform the requested operation after preflight. Installation and model downloads are side effects and must be limited to the selected project.

Route a request

1. Identify the objective: domain adaptation (continued pretraining), examples-to-behavior (SFT), preferences (DPO/ORPO/KTO), reward-driven behavior (GRPO/RL), embeddings, vision, speech, or inference/export.
2. Identify the adaptation method independently: full fine-tuning, LoRA, or QLoRA.
3. Inspect model family, revision, modality, chat template, dataset schema, sequence length, GPU memory, and desired output format.
4. For Discuss, use the references without running commands. For Prepare, create a JSON config and conversion script. For Execute, run the command below from this skill directory.

Runner