Lists multiple specific concrete actions: SSH job queue, multi-seed/multi-config ML experiments, OOM-aware retry, stale-screen cleanup, and wave-transition race prevention. These are highly specific capabilities.

Clearly answers both 'what' (SSH job queue for ML experiments with OOM-aware retry, stale-screen cleanup, wave-transition race prevention) and 'when' (explicit 'Use when' clause with specific trigger phrases and a clear threshold of '10+ jobs that need orchestration').

Excellent coverage of natural trigger terms including 'batch experiments', '队列实验' (Chinese), 'run grid', 'multi-seed sweep', 'auto-chain experiments', and even references the boundary with '/run-experiment' for 10+ jobs. These are terms users would naturally say.

Highly distinctive with a very specific niche: SSH-based ML experiment orchestration with particular features like OOM-aware retry and wave-transition race prevention. The explicit boundary with '/run-experiment' for smaller jobs further reduces conflict risk.

The skill is quite long (~300 lines) and includes sections like 'Why This Exists' with rationale/post-mortem context, a comparison table with /run-experiment, and 'Rationale / Source' that don't add actionable value. The 'Known Failure Modes' and core workflow sections are efficient, but overall there's significant content that could be trimmed—Claude doesn't need the backstory of a 2026-04-16 audit or explanations of why engineering friction exists.

The skill provides concrete YAML manifest formats, regex patterns for OOM detection, and SSH commands for launching/monitoring. However, the actual scheduler implementation (queue_manager.py) is referenced as 'bundled' but not provided, and no bundle files exist. The workflow relies heavily on tools that aren't present, making it not truly executable. The grid spec and wave chaining examples are well-structured but remain declarative specifications rather than runnable code.

The 5-step workflow (Parse → Pre-flight → Launch → Monitor → Post-completion) is clearly sequenced with explicit validation at each stage. The job state machine is well-defined with transitions, the wave transition has 4 explicit preconditions before proceeding, OOM handling has a clear retry loop with bounded attempts, and stale screen detection has explicit decision logic. The resume-on-restart section addresses crash recovery.

The skill references bundled tools (queue_manager.py, build_manifest.py) and related skills (/run-experiment, /monitor-experiment, /analyze-results) but no bundle files are provided. The content is largely monolithic—the YAML manifest format, grid spec syntax, wave chaining, OOM handling, and stale screen detection could be split into separate reference files. The 'See Also' section is good but the main body tries to cover everything inline.