How to use the ML Training Recipes agent skill

What the skill makes the agent do

An agent skill is a set of instructions and reference files that the agent loads when a task matches the skill's description. With ml-training-recipes installed, the agent follows the skill's established practices when you ask about training. The source describes patterns drawn from production codebases and modern training practice, covering training loops, optimizer selection, LR scheduling, mixed precision, debugging, and systematic experimentation.

The skill is organised around a few reference files. SKILL.md holds the core guidance on architecture selection, scaling, training loops, optimizers, schedules, mixed precision, memory and hyperparameter search. Other files go deeper: architecture.md (transformer code patterns and weight initialisation), optimizers.md (Muon, AdamW hybrid, per-group learning rates), domain-specific.md (vision, diffusion, distributed training, checkpointing, data loading), scaling-and-selection.md (scaling laws and compute budgets), biomedical.md (drug discovery, protein models, medical imaging, genomics) and experiment-loop.md (an autonomous keep/discard/revert experiment loop).

• Architecture choice by data type and sample count, for example pretrained CNN with fine-tuning for images under 10K samples.
• Optimizer defaults such as AdamW with beta1 0.9, beta2 0.95, and eps 1e-10 for bf16 training.
• Gradient clipping at 1.0 for Transformers, with the note that Muon may not need it.
• OOM fixes listed in order, starting with reducing the device batch size and raising gradient accumulation.
• Loss-spike and divergence checks, including a fast-fail when loss exceeds 100.

Install the skill from the ZIP

Download the ml-training-recipes ZIP from this page and unzip it. You can install it for yourself or for a single project.

For a personal install, unzip the folder into ~/.claude/skills/. For a project-only install, unzip it into the project's .claude/skills/ folder. Either way the skill folder should contain SKILL.md and the references directory.

If you use the Claude apps instead, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal: ~/.claude/skills/
• Project: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills

Check the requirements before you start

The skill assumes a working Python environment with PyTorch. The metadata lists torch>=2.0.0 as the dependency. An NVIDIA GPU is recommended because the bf16, TF32 and torch.compile guidance targets NVIDIA hardware such as Ampere-class and newer GPUs.

The skill does not run scripts and does not use the network, according to its listing. It is guidance the agent reads and applies to your code.

• Python with PyTorch, torch>=2.0.0
• NVIDIA GPU recommended for bf16, TF32 and torch.compile advice
• Your own training code and data; the skill does not supply a dataset

Trigger the skill with natural requests

You do not need a special command. The skill is meant to load when your request matches its description: training or fine-tuning a network, debugging loss spikes or out-of-memory errors, choosing an architecture, or improving GPU throughput. Describe your problem in plain language and include the details that matter, such as your data size, GPU and symptoms.

Good requests name the situation. Vague requests like 'make my model better' give the agent less to work with.

• 'I am fine-tuning a ViT on 8,000 labelled images. Which model and learning rate should I start with?'
• 'My loss jumps to NaN after about 2,000 steps. Walk me through how to diagnose it.'
• 'I get CUDA out of memory when batch size is 32. What should I change first?'
• 'Check my training loop for mixed precision and gradient clipping problems.'
• 'My GPU utilisation is low during training. Help me find where the time goes.'

Examples of what you get back

For an OOM question, the skill's ordered list suggests trying a smaller device batch size with more gradient accumulation first, then setting PYTORCH_ALLOC_CONF to expandable_segments:True, then zeroing gradients with set_to_none=True. Later options include meta-device initialisation, activation checkpointing and an 8-bit optimizer through bitsandbytes.

For a training-loop review, the agent can check that gradients are clipped, that TF32 matmul precision and bf16 autocast are used on supported GPUs, and that the first steps are excluded from timing because of JIT compilation.

Safety, limits and review

The recipes are general patterns. The skill's own description warns that the agent may apply hyperparameters that do not suit your data or hardware. Treat its suggestions as a starting checklist, not a verified configuration.

Because the skill can suggest learning rates, weight decay and schedules, check them against your validation results. For production or high-stakes models, have an experienced reviewer look at the training setup. The skill includes a fast-fail check that exits when loss exceeds 100, which you should adjust to your own loss scale.

• Verify suggested hyperparameters on a small run before a long job.
• Review any code the agent edits, especially training loops and data handling.
• Expect the guidance to assume PyTorch and GPU experience.

Tips for better results

Give the agent your hardware, data size, model family and current metrics in the first message. The skill's architecture and scaling tables depend on these. If the agent's answer ignores a detail, point to it and ask it to re-check against the relevant reference file.

Ask for the specific reference file when you need depth, for example biomedical.md for protein or genomics models, or experiment-loop.md for a systematic keep/discard experiment process.

• State GPU model and memory, data type and sample count.
• Paste the exact error or loss curve, not a paraphrase.
• Ask one question at a time when debugging so each change can be tested.