The body is dense with tables and executable snippets and never explains basics Claude already knows; deep material is offloaded to references, so each section is lean despite the multi-domain scope.

Provides copy-paste-ready code for the training loop, LR schedules, and compile setup plus concrete optimizer/LR tables and an ordered OOM remedy list, matching the executable-and-specific anchor.

The training loop includes an explicit divergence fast-fail checkpoint ('if loss.item() > 100'), and the debugging/OOM/plateau sections are ordered checklists with diagnostic-to-fix feedback, meeting the clear-sequence-with-checkpoints anchor.

A 'Reference files (read when needed)' section lists six real one-level-deep reference files, with inline pointers (e.g., 'For biomedical domains, see references/biomedical.md'), keeping the body an overview and details one hop away.

Lists multiple concrete actions — 'training loops, optimizer selection (AdamW, Muon), LR scheduling, mixed precision, debugging, and systematic experimentation' — matching the anchor for listing several specific capabilities rather than vague domain-only phrasing.

Explicitly answers both what ('Battle-tested PyTorch training recipes for all domains…') and when (an explicit 'Use when…' clause), satisfying the top anchor rather than capping at 2 for a missing trigger.

'Use when training or fine-tuning neural networks, debugging loss spikes or OOM, choosing architectures, or optimizing GPU throughput' covers natural phrases a user would actually say, spanning fine-tuning, loss spikes, OOM, and throughput.

The PyTorch training-recipe niche with domain list and specific triggers (loss spikes, OOM, GPU throughput) is clearly distinguishable and unlikely to fire for unrelated skills.