Content
100%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.
The body is a lean, fully executable guide: a hardware policy, a gated five-step workflow, copy-paste benchmark and ncu commands, and clear deferral of correctness evidence and sustained iteration to linked docs. No dimension shows more than minor room for improvement.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~48-line body assumes Claude's competence throughout — no explanations of what Nsight Compute or benchmarking are — and every sentence earns its place with project-specific constraints or commands (e.g. "The benchmark runner launches a child process, so use `--target-processes all`"). | 5 / 5 |
Actionability | Commands are copy-paste ready with concrete flags and arguments (`FLA_BENCH_BASE=$(git rev-parse origin/main)`, `python benchmarks/benchmark_training_throughput.py --name kda --batch_size 2 --seq_len 8192 --varlen`, full `ncu` invocations), and the one needed substitution is explicitly flagged ("Replace the kernel filter with the changed kernel's name"). | 5 / 5 |
Workflow Clarity | The five-step workflow has explicit validation gates ("Pass the relevant correctness tests before timing a candidate", "rerun correctness and timing", "does not replace the full relevant gate before reporting a gain") and a measure-optimize-rerun feedback loop, so it matches the anchor with checkpoints rather than falling to 4. | 5 / 5 |
Progressive Disclosure | At under 50 lines with no bundle files needed, the four well-organized sections plus clearly signaled one-level links to CONTRIBUTING.md#benchmarking and the fla-optimization-loop skill meet the rubric's simple-skill exception for a top score. | 5 / 5 |
Total | 20 / 20 Passed |