Why use it: you want to fix the tool, not use it — patch scripts, fix docs, or propose methodology changes. What it produces: a compliant PR, or an issue accepted by the maintainer.
Command SHELL + GitHub operations. This track is independent of 01–07.
Rule number one, read before typing
SKILL.md is a core asset and does NOT accept external PRs. Every line in it has been verified and optimized according to the maintainer's version. Find a bug or improvement idea in the methodology → open an issue, accepted ideas will be written in by the maintainer and you'll be credited in the commit (precedent: PR #59 found an out-of-bounds bug description).
| Contribution type | Path |
|---|---|
| Methodology bug/idea | open an issue, don't PR directly into SKILL.md |
Fix a script in scripts/ |
direct PR, with reproduction steps |
| Translate README / fix typos | direct PR |
| Your persona skill | doesn't go into examples/ — see 07 |
| Derivative tools / collections / orchestration | PR adding a line to COMMUNITY.md |
Set up the environment
No build step, no package manager, no requirements.txt. Just Python 3 and (for one script) yt-dlp:
git clone https://github.com/alchaincyf/nuwa-skill && cd nuwa-skill
python3 --version # 3.9+ (code dùng type hint list[str])
command -v yt-dlp || brew install yt-dlp
Run all existing checks
The repo has no test suite. The real verification loop is running the script on the available examples:
# Cấu trúc research 01-06 (Jobs, Karpathy, Trump, MrBeast, PG, Ilya, Zhang Yiming, Zhang Xuefeng, Sun Yuchen)
python3 scripts/merge_research.py examples/steve-jobs-perspective # kỳ vọng thoát 0, in bảng
# Cấu trúc references phẳng (Munger, Feynman, Taleb, Musk, Naval)
python3 scripts/merge_research.py examples/munger-perspective # kỳ vọng thoát 1: 目录不存在
python3 scripts/quality_check.py examples/steve-jobs-perspective/SKILL.md # kỳ vọng 6/6, thoát 0
python3 scripts/quality_check.py examples/munger-perspective/SKILL.md # kỳ vọng 5/6, thoát 1
printf '1\n00:00:01,000 --> 00:00:03,000\n<i>Hello</i> world.\n\n2\n00:00:03,000 --> 00:00:05,000\nHello world.\n' > /tmp/demo.srt
python3 scripts/srt_to_transcript.py /tmp/demo.srt # kỳ vọng gỡ thẻ <i> và khử dòng lặp
Scan all 15 examples to see where the script stands before and after your fix:
for d in examples/*/; do
printf '%-40s ' "$d"
python3 scripts/quality_check.py "$d/SKILL.md" 2>/dev/null | grep '结果:' || echo "LỖI"
done
Machine check for community intake PRs
GITHUB_TOKEN=$(gh auth token) python3 .github/scripts/community_check.py --check-repo <owner>/<repo>
This mode is exactly the manual test mode that the file itself documents in its docstring. CI mode (without --check-repo) needs the variables GITHUB_TOKEN, GITHUB_REPOSITORY, PR_NUMBER and will post a comment on the PR — don't run that mode locally.
Safety note when modifying the workflow: it uses pull_request_target (runs with write permission on the origin repo) and intentionally only checks out the script on main, never checks out or runs PR code. Keep that property; if you add a PR checkout step, that's a privilege escalation vulnerability, not an improvement.
Checklist before opening a PR
- Don't touch
SKILL.md - No
.DS_Storesimilar junk branch - One PR does only one thing (don't bundle multiple characters into one PR)
- The PR description states clearly: what it does, why, and how it was verified
Open tasks, if you want to find somewhere to contribute
download_subtitles.shhas a dead branch: the Chinese subtitle branch detects new files usingfind ... -newer /tmp/.ytdlp_markerwhile/tmp/.ytdlp_markeris not created anywhere in the repo, so that branch never reports success and always falls through to the English branch (costing an extra call toyt-dlp). Branches 2 and 3 use-mmin -1so they still find the file; fixing branch 1 to be consistent is a tidy PR, exactly the kind of "script fix" that gets accepted directly.merge_research.pycan only readreferences/research/01-06.md, while 5/15 of the project's own examples use flat files with freely chosen names — either extend the script, or standardize the example structure.quality_check.pycount 诚实边界 with regex list item, so an honest boundary written as a paragraph or table gets scored 0 (munger-perspectiveitself has this bug).