review-nvbench
Skill945 starsChanged 5 months agoApache-2.0View on GitHub
Use this skill when the user invokes /review-nvbench to review an NVBench pull request, PR branch, commit range, or diff. Supports context, feedback, and adversarial modes.
In this repositoryOne other file it gives its agents.
- AGENTS.mdAGENTS.md
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 5 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 8
Read AGENTS.md and docs/pr_review.md before acting.
- Distinctive: line 10
Use one of these modes:
- Very specific: line 16
If no mode is provided, default to context.
- Very specific: line 20
Run or inspect ci/util/pr_review_context.sh when reviewing a checked-out PR branch.
- Distinctive: line 30
Do not produce review findings in context mode.
- Distinctive: line 34
Before producing findings, make sure the context pass has happened.
- Distinctive: line 42
Before producing adversarial feedback, make sure the PR intent and implementation strategy are understood.
- Distinctive: line 51
Keep this distinct from normal feedback.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Review NVBench Pull Request
Read AGENTS.md and docs/pr_review.md before acting.
Use one of these modes:
context: Gather PR context and explain the change without review findings.feedback: Produce normal review findings after the human reviewer has inspected the PR.adversarial: Challenge assumptions, design direction, compatibility risks, and hidden failure modes after the PR context is understood.
If no mode is provided, default to context.
Context Mode
Run or inspect ci/util/pr_review_context.sh when reviewing a checked-out PR branch. If the user provided a PR number, pass --pr <number>. If the user provided an issue number, pass --issue <number>.
Summarize:
- The linked issue or PR context.
- The problem the PR is trying to solve.
- How the PR attempts to solve it.
- The important changed files and a sensible human review order.
- Any missing context, such as no discoverable linked issue.
Do not produce review findings in context mode.
Feedback Mode
Before producing findings, make sure the context pass has happened. If it has not, gather and summarize the PR context first.
Review for correctness, benchmark validity, performance, API compatibility, build and packaging impact, and consistency with existing code. Let the changed files determine which risks matter, while paying attention to relevant NVBench contracts such as CUDA stream ordering, timing boundaries, cold/batch/CPU-only measurement behavior, manual timer semantics, axis generation, output schemas, optional CUPTI/NVML behavior, CMake options, Python bindings, and test coverage.
Report findings first, ordered by severity, with file and line references. Do not post comments to GitHub; the human reviewer decides which findings to use.
Adversarial Mode
Before producing adversarial feedback, make sure the PR intent and implementation strategy are understood.
Focus on whether the approach is sound:
- Hidden assumptions that could fail.
- Simpler or safer alternative designs.
- Compatibility, layering, maintenance, or rollout risks.
- Cases where the implementation solves the local issue but weakens a broader NVBench contract.
Keep this distinct from normal feedback. It is acceptable for adversarial mode to produce questions or risks rather than concrete blocking findings.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

