| name | cuda-review-helper |
| description | Review CUDA kernels, progression folders, Makefiles, and interactive visualizations in cuda-algorithm-lab with a focus on correctness, readability, educational flow, and repo consistency. |
CUDA Review Helper
Use this skill when the user wants a review of CUDA code, a folder progression,
an HTML visualization, or overall repository quality in this repository.
Review Priorities
-
CUDA correctness
- Check indexing logic carefully.
- Check boundary conditions and power-of-two assumptions.
- Check synchronization placement such as
__syncthreads().
- Check shared-memory usage, bank-conflict comments, and any mismatch
between comments and implementation.
- Check whether kernels appear internally consistent with their launch
configuration.
-
Educational clarity
- Prefer files that explain the algorithm at the top in plain language.
- Look for progression quality: each numbered step should justify why it
exists and how it improves on the previous step.
- Check whether comments explain intent rather than restating syntax.
- For visualization files, check that the UI actually teaches the same
algorithm the CUDA file implements.
-
Repository consistency
- Folder numbering should match filenames, README references, and Makefiles.
- Root-level naming should avoid duplicate or superseded examples.
- Local plugin metadata, repo URLs, and marketplace entries should match the
current repository identity.
-
Submission readiness
- Call out broken Makefile paths, stale references, temp files, unfinished
code, or obvious compile-time issues.
- Prefer actionable findings over broad summaries.
Workflow
- Read the target files and nearby README/Makefile context.
- If reviewing code, compare comments with actual implementation behavior.
- If reviewing a visualization, compare the visual explanation with the CUDA
kernel it is supposed to teach.
- Report findings ordered by severity.
- If there are no findings, say so explicitly and mention any residual risks,
such as "not compiled in this session."
Output Style
- Findings first, with file references.
- Keep summaries brief.
- Focus on bugs, regressions, stale docs, misleading explanations, and missing
educational context before style-only suggestions.