使用方法
- 复制安装命令,在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物,并对照来源检查输出。
npx skills add https://github.com/vincenzoimp/academic-research-skills --skill develop-contribution开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况,不等于独立质量或安全测试。
按源仓库内容呈现,保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Develop Contribution
Create or regularize one contributions/<slug>/ folder: badge-general compliant, self-contained, with a report/report.tex detailed enough that paper writing never needs to re-read the code.
Read First
contributions/README.md— the badge-general contractsurvey/survey.pdf, gaps section — when positioning a new contributionreferences/methodology.md— evaluation, statistics, figure disciplinereferences/experiments.md— run logs and autonomous campaignsreferences/reproduction.md— reproducing external workreferences/ethics.md— data red flagsreferences/data.md— data placement and inter-contribution data
dependencies (when the contribution has shared/large data or consumes another contribution's output)
Before Building (both modes)
Fill the Positioning section of the contribution README:
- the claim this contribution will support
- the delta vs the nearest prior work (cite survey/SOTA citekeys)
- the evidence plan: what will be measured or collected, and how
- the falsifiability condition: what outcome would refute the claim
If the claim has no evidence path, stop and discuss with the user.
Create Mode
- Copy
contributions/_template/tocontributions/<slug>/(kebab-case
slug).
- Python code? Write the contribution's own
pyproject.toml(name,
requires-python, dependencies), add "contributions/<slug>" to [tool.uv.workspace] members in the root pyproject.toml, run uv sync --all-packages from the root (plain uv sync prunes the members' dependencies on a package = false root). Genuinely conflicting dependencies → move the path to exclude, create a local venv, document it in the README.
- Develop inside the folder — free-form structure (
src/,data/,
figures/, outputs/, notebooks). Follow references/methodology.md for any evaluation; keep a run log per references/experiments.md for every run that supports a claim, including autonomous campaigns. For reproductions of external work, follow references/reproduction.md.
- Fill the README badge checklist truthfully — every checked box has
evidence inside the folder. Check data against references/ethics.md.
- Write
report/report.tex: motivation, positioning vs the SOTA (cite the
root .bib), method, setup, results with figures/tables, limitations. Build it: make contribution C=<slug>.
- Run
make check.
Regularize Mode
- Inventory the draft folder: what exists, what it claims, what is missing
against the badge checklist.
- Reorganize in place — keep provenance (note original filenames when
renaming), capture the environment, verify the run path works end-to-end, record data provenance and ethics flags.
- Continue with Create Mode steps 4–6.
Rules
- Self-contained: no imports from other contributions; tooling worth
sharing becomes its own contribution referenced in the README. If the input is another contribution's output, declare it and read it in place — never copy (see references/data.md).
- Exploratory results never masquerade as confirmed evidence — label them.
- Negative results are results: report them in report.tex.
Done When
- Badge checklist fully and truthfully checked,
report.pdfbuilt,
workspace coherent, make check passes


