What does this skill do, and when should you use it?
This skill analyzes large NVIDIA Megatron-LM pull requests and proposes a split based on CODEOWNERS reviewer groups. It keeps tests with the production code they validate, identifies dependencies between resulting PRs, and requires user approval before execution. After approval, it creates branches from appropriate bases, applies file-scoped diffs, pushes to the user’s fork, and opens draft PRs.
Fetches PR title, body, source branch, author, diff statistics, and the current GitHub user; parses .github/CODEOWNERS and maps changed files to reviewer groups; clusters files by ownership and counts the PR’s distinct groups; produces a split plan with files, required groups, and dependencies; after approval, uses git diff and git apply to extract file-scoped changes, commits and pushes them to the user’s fork, and creates draft PRs.
- A Megatron-LM contributor needs to divide a large PR spanning several ownership areas into smaller reviewable PRs.
- A PR author wants to reduce the number of CODEOWNERS reviewer groups involved in each PR.
- A maintainer needs tests to remain with the production code they validate instead of becoming a separate test-only PR.
- A change includes renamed symbols or cross-PR dependencies that require compatibility aliases, re-exports, or shims.
How do you install this skill?
- Execution reads PR data, CODEOWNERS, and the current GitHub user, and may create branches, commits, pushes, and draft PRs; confirm the repository, bases, file scope, and fork target before execution.
- There is no defined recovery procedure for gh authentication failure, git apply conflicts, push failure, partially created PRs, or rollback.
- BENCHMARK.md states that Tier 3 was unavailable and evals.json is empty; it also records missing examples, Instructions, metadata.tags, and an explicit WHEN-to-use description.
- The skill depends on GitHub CLI, Git, upstream/main, and NVIDIA Megatron-LM repository conventions; operation on other hosting platforms or through mainland-China networks is not established.
- Shell / CLI
- Network access
- Local filesystem
GitHub CLIGit
Install the specific skill with the default skills CLI:
npx skills add nvidia/skills --skill mcore-split-pr --yes
The README does not specify a client-specific local folder; the CLI prompts for the installation destination.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/mcore-split-pr" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Split PR https://github.com/NVIDIA/Megatron-LM/pull/123 to reduce CODEOWNERS reviewers
- Split PR 123 into independently mergeable draft PRs.
After installation, ask the agent for example: “Split PR https://github.com/NVIDIA/Megatron-LM/pull/123 to reduce CODEOWNERS reviewers” or “Split PR 123 into independently mergeable draft PRs.” The agent should present the constraints and proposed split first, then wait for approval before executing.
What are this skill's strengths and limitations?
- Optimizes around CODEOWNERS groups rather than file count alone.
- Explicitly preserves production-test relationships and checks independent mergeability.
- Requires pushes to the user’s fork and draft PR creation during execution.
- Its scope is specifically NVIDIA/Megatron-LM pull requests and their CODEOWNERS file.
- Execution requires GitHub CLI, Git, network access, and permission to use the user’s fork.
- The SKILL.md provides no test suite, platform matrix, or failure-recovery procedure.
- The user must approve the proposed split before execution.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| CODEOWNERS PR Splitter this page ✓ NVIDIA · Official | 57 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| Codex Pull Request Editor ✓ OpenAI · Official | 50 · Use with care | ★ 129k | 13d ago | Apache-2.0 |
| Yeet GitHub PR Flow ✓ OpenAI · Official | 41 · Not recommended | ★ 28k | 4mo ago | — |
| CodeRabbit Autofix | 53 · Use with care | ★ 193 | 12d ago | MIT |
| Ship a PR (e2e repo PR delivery workflow) | 59 · Recommended | ★ 10k | 13d ago | Apache-2.0 |
How did FollowSkills review this skill?
The skill requires user approval before execution, creates draft PRs, pushes to the user's fork rather than upstream, and exposes the main git/GitHub operations, reducing unauthorized and irreversible-action risk. It does not explain credential handling, sensitive-code treatment, rollback, or remote-branch cleanup, so full marks are not justified.
The workflow and key commands are broadly self-consistent and cover CODEOWNERS analysis, dependency identification, and split execution. However, it depends on gh, git, upstream/main, and repository configuration, with little guidance for invalid input, command failure, conflicts, recovery, or diagnosable error feedback; static calibration also caps this dimension at 10.
Trigger phrases, input parameter, audience, and primary scenario are reasonably clear for developers splitting Megatron-LM PRs by CODEOWNERS groups. Boundaries for unsuitable cases, complex dependencies, or non-GitHub environments are underdeveloped, and Chinese-language support and mainland-China network reachability are not addressed.
SKILL.md provides front matter, layered workflow guidance, constraints, and operational safeguards; Apache-2.0 licensing, author, version, and benchmark information are present. However, metadata.tags, Instructions, Examples, FAQs, detailed limitations, changelog, and explicit maintenance responsibility are missing, matching the benchmark findings.
The skill defines a useful path from PR inspection and CODEOWNERS clustering through draft-PR creation, including test co-location and dependency handling. There is no representative directly usable output or execution reproduction for this skill path, and results still require user review and repository-specific handling; static calibration therefore limits the score.
A committed static benchmark reports nine Tier 1 checks with ten findings and states that Tier 3 was unavailable. evals.json is empty, and there are no dedicated key-path tests or independent reproduction materials, so only limited evidence is supportable.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →