Content
82%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A lean, expert-oriented reference with genuinely actionable config values, env vars, and measured numbers, plus thoughtful epistemic hygiene (rounded bands, anti-cargo-culting warnings, A/B causality caveats). The main gaps are redundant repetition of the same parallelism shapes across three sections and top-of-file references to files that are not part of the skill bundle.
Suggestions
Consolidate the TP/EP/PP/VPP shapes so they appear authoritatively in one table (e.g., the First Answer Checklist) and are referenced rather than restated in the Rounded Performance Bands and Config Families sections.
Fix the garbled run-on sentence in the First Answer Checklist ('Treat TE-scoped CUDA graph scopes ... as profile-driven candidates, CUDA_DEVICE_MAX_CONNECTIONS selection, ... and the warning not to cargo-cult tracker rows.') by recasting it as a proper list of tuning knobs.
Either include the referenced files (docs/training/moe-optimization.md, card.yaml) in the skill bundle or remove/qualify the top-of-file references so navigation does not dead-end.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense, tabular, and assumes Claude's expertise — it explains nothing the model already knows, and the symbol legend ('E = embedding ... | = stage boundary') is only as long as needed. It sits at 4 rather than 5 because the TP/EP/PP/VPP shapes are repeated across the First Answer Checklist, Rounded Performance Bands, and Representative Config Families sections, which could be consolidated. | 4 / 5 |
Actionability | The content is fully concrete: exact configs ('TP=2 EP=64 PP=8 VPP=4', 'Dispatcher: DeepEP'), copy-paste environment variables (CUDA_DEVICE_MAX_CONNECTIONS=1, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, NCCL_GRAPH_REGISTER=0), and measured results ('20.14729s/step, 299.352 model TFLOPS/GPU'). Hedges such as 'VPP=2 or similar' and the rounded bands are explicitly justified ('intentionally rounded so the document stays durable as the tracker moves'), which the rubric permits. | 5 / 5 |
Workflow Clarity | Usage guidance is clear and ordered ('answer from these canonical rows before adding throughput caveats'; 'explicitly say VPP=unspecified; do not invent'), and a validation methodology is explicitly stated in Pitfall 7 (short screening runs, matched one-variable A/Bs, a 50-step final run). It is not a 5 because this methodology is not laid out as an explicit sequenced process with checkpoints in the main flow, and not a 3 because validation and causality separation are concretely described. | 4 / 5 |
Progressive Disclosure | The body has good structure — clear sections, tables, and top-of-file references that are clearly signaled ('Stable docs: @docs/training/moe-optimization.md', 'Card: @skills/nemo-mbridge-perf-moe-hardware-configs/card.yaml'). However, no bundle files exist in the skill (no references/, scripts/, or assets/), the referenced paths point outside the skill, and all ~200 lines of reference tables are inlined in a single file, keeping it below anchor 5. | 4 / 5 |
Total | 17 / 20 Passed |