Content
75%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 well-structured, actionable profiling workflow with concrete tools and a tracked instrumentation changelog; main improvements would be adding an explicit post-instrumentation validation checkpoint and optional code snippets.
Suggestions
Add an explicit validation checkpoint in Step 5 (e.g. 'Run the target once after instrumentation to confirm it still executes correctly before collecting profiling data').
Include one or two short, copy-paste instrumentation snippets (e.g. a CUDA event timing context manager or an NCCL collective wrapper) to lift actionability from concrete commands to fully executable guidance.
Consider moving the per-tool reference lists into a references/ file and linking from SKILL.md to reduce inline length and improve progressive disclosure.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly lean: tool lists and measurement dimensions are given without explaining concepts Claude already knows, though a few explanatory phrasings (e.g. 'Don't limit yourself to the examples below') could be trimmed. | 4 / 5 |
Actionability | Provides concrete commands (nvidia-smi dmon, nvidia-smi topo -m, NCCL_DEBUG=INFO, strace -c, perf stat) and specific instrumentation scenarios, but lacks copy-paste instrumentation code examples, leaving minor gaps. | 4 / 5 |
Workflow Clarity | Six clearly sequenced steps with a clarifying 'ask the user' loop and a MANDATORY instrumentation changelog plus cleanup offer; the main gap is no explicit checkpoint verifying the instrumented program still runs before profiling. | 4 / 5 |
Progressive Disclosure | Well-organized into clearly headed steps with no nested references and all content self-contained; slightly over 50 lines with tool reference lists that could optionally be split into a reference file, so not a 5. | 4 / 5 |
Total | 16 / 20 Passed |