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 strong, highly actionable body: exact config snippets, exact commands, explicit constraints, pitfalls with failure modes, and a verification section with concrete success criteria. The main gaps are minor — some trimmable quoted upstream code, no recovery loop after the "HCP is not active" failure signal, and code anchors that could be pushed to a reference file.
Suggestions
Add a fix-and-retry step after the "If you only see CONTEXT_PARALLEL_GROUP, HCP is not active" check (e.g., confirm hierarchical_context_parallel_sizes was passed through the recipe and that use_decentralized_pg is False, then relaunch) to close the workflow feedback loop.
Move the verbatim upstream docstring and initialization.py excerpts into a single reference file, keeping only the line-anchored paths in SKILL.md, to tighten the body and improve progressive disclosure.
Drop the duplicate statement of the product constraint (Pitfall 5 restates "prod(hierarchical_context_parallel_sizes) == context_parallel_size" from Required constraints) or merge them into one place.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational throughout — no basic-concept padding, every section (Enablement, Code Anchors, Pitfalls, Verification) earns its tokens. Minor trim opportunities exist, such as the 8-line upstream docstring quote in model_parallel_config.py and the slightly redundant restatement of the product constraint in both "Required constraints" and Pitfall 5, keeping it just short of fully lean. | 4 / 5 |
Actionability | The skill gives copy-paste-ready guidance end to end: an exact Bridge override snippet (context_parallel_size=4, cp_comm_type="a2a+p2p", hierarchical_context_parallel_sizes=[2,2]), exact runnable commands (uv run pytest on a named test file; a fully specified 4-GPU torch.distributed.run invocation with recipe and overrides), and explicit success criteria. | 5 / 5 |
Workflow Clarity | There is a coherent arc from enablement through constraints to pitfalls to a Verification section with explicit validation checkpoints (log check for HIERARCHICAL_CONTEXT_PARALLEL_GROUPS, "Training completes at least one step") and a failure signal ("If you only see CONTEXT_PARALLEL_GROUP, HCP is not active"). It stops short of level 5 because the failure signal has no corresponding fix-and-retry loop guiding recovery when HCP is not active. | 4 / 5 |
Progressive Disclosure | Structure is good: short body, clear section headers, and well-signaled one-level-deep references up front ("For what hierarchical CP is, when to use it, and the decision tree..., see: @docs/... and @skills/.../card.yaml"). It does not reach 5 because the inlined upstream code-anchor excerpts (docstrings and initialization.py fragments) are detail that could live in a separate reference file, and there is no reference organization beyond the two intro links. | 4 / 5 |
Total | 17 / 20 Passed |