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 retry, cleanup, race prevention) and 'when' (explicit 'Use when' clause with specific trigger phrases and a clear threshold of 10+ jobs needing 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 niche: SSH-based ML experiment orchestration with specific features like OOM-aware retry and wave-transition race prevention. Explicitly distinguishes itself from '/run-experiment' for smaller jobs, reducing conflict risk.

The skill is quite long (~350+ lines) and includes some sections that could be trimmed—e.g., the 'Why This Exists' rationale, the 'Rationale / Source' section, the comparison table with /run-experiment, and detailed explanations of legacy shim chains. However, most content is genuinely instructive and not explaining things Claude already knows (it's project-specific orchestration logic). The legacy resolver chain in Step 3a is particularly verbose for what it accomplishes.

The skill provides fully executable bash commands, concrete YAML manifest examples, specific regex patterns for OOM detection, jq commands for monitoring, and complete code blocks for every step from manifest creation through launch, monitoring, and resume. The grid spec syntax and wave chaining examples are copy-paste ready with clear variable substitution patterns.

The workflow is clearly sequenced (Steps 1-5) with explicit validation checkpoints: pre-flight checks in Step 2, state machine with clear transitions, wave transition conditions (4 explicit criteria), stale screen detection logic, OOM retry with bounded attempts, and resume-on-restart procedures. The feedback loops for error recovery (OOM → retry, stale → cleanup → requeue) are well-defined with explicit conditions.

The skill references external scripts (queue_manager.py, build_manifest.py) and other skills (/run-experiment, /monitor-experiment, /analyze-results) appropriately, and the 'See Also' section provides clear navigation. However, the SKILL.md itself is monolithic—sections like the detailed resolver chain, the comparison table, the rationale section, and the known failure modes could be split into separate reference files. No bundle files were provided to verify referenced paths exist.