Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add phanghonghao/THU-Awesome-Skills --skill paper-reprogit clone --depth 1 https://github.com/phanghonghao/THU-Awesome-SkillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/phanghonghao/thu-awesome-skills/paper-repro)<a href="https://agentmods.dev/skills/phanghonghao/thu-awesome-skills/paper-repro"><img src="https://agentmods.dev/badge/skills/phanghonghao/thu-awesome-skills/paper-repro/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/phanghonghao/thu-awesome-skills/paper-repro"><img src="https://agentmods.dev/badge/skills/phanghonghao/thu-awesome-skills/paper-repro.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00146 | $0.02180 |
| Opus 5.5 | $0.00058 | $0.00872 |
| Sonnet 5.5 | $0.00029 | $0.00436 |
| Haiku 4.5 | $0.00015 | $0.00218 |
Grade B, and why
paper-repro scanned grade B with 2 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Asks for rootmediumPrivilege escalation
A mod that escalates privileges can change anything on the machine, not only the project.
- **禁止 sudo / 管理员权限**,禁止改系统环境变量、禁止写用户主目录之外。 Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
所有脚本网络层均为 **requests → curl → 警告退出**,对齐 `/web-search-fallback`,MCP 限流时仍可用。 How it starts
The opening of the file, as written. The whole thing — 138 lines — stays where its author put it; the contents beside it link to each section on GitHub.
paper-repro — 论文「总结 + 最小复现」
把一篇论文变成:一页可读总结 + 一个能在本机跑通的最小复现 + 完整复现成本估算。
编排 /paper-html-onepage(读懂)+ 本 skill 的复现脚本(跑通)。
⚠️ 安全约束(自动运行模式必备,优先于一切)
- 只跑最小规模:限定步数/epoch(默认 ≤ 几百步)、限定时长(默认 ≤ 几分钟)。
- 隔离运行:所有复现代码、数据、日志放进单独子目录,不污染其他项目。
- 禁止 sudo / 管理员权限,禁止改系统环境变量、禁止写用户主目录之外。
- 网络三思:允许浅克隆(
git clone --depth 1)和小数据下载;超过 ~200MB 的下载、需登录/付费的数据集、需 Kaggle/HF token 的资源,先问用户,不要自动拉取。 - 不确定就不跑:如果生成的代码看不清在干嘛、或需重型依赖(CUDA 专版、编译、root),只产出方案+脚本,不自动执行,把决定权交给用户。
输入解析(Stage 0)
接受四种输入,统一抽出 arXiv ID + 标题 + 输出目录:
- arXiv URL:
https://arxiv.org/abs/1706.03762或/pdf/...→ 抽 ID - arXiv ID:
1706.03762 - 关键词/标题:如
Attention Is All You Need - 本地 PDF:
--pdf <path>(复现阶段需要标题,从文件名/首页推断)
输出目录默认 <用户当前工作目录>/<论文短名>_repro/。
6 阶段编排流程
本 skill 是 agent 驱动 的:脚本(
scripts/)只做查找/分析/解析等机械活, 论文级的代码改写由你(agent)按方法论完成——因为每篇论文不同,无法写成单一固定脚本。 这与/web-search-fallback「指令 + 可复用片段」的设计哲学一致。
Stage A — 读懂论文(委派 /paper-html-onepage,不改它)
python "<SKILL_ROOT>\paper-html-onepage\scripts\paper_to_onepage_html.py" \
--url "<arxiv_url>" --out "<outdir>/<name>_summary.html"
(关键词搜用 --query;本地 PDF 用 --pdf。)得到 <name>_summary.html。
Stage B — 找代码(scripts/find_repo.py)
python "<SKILL_ROOT>\scripts\find_repo.py" "1706.03762" --top 6
# 加 --json 拿结构化结果
arXiv ID → 自动取标题 → GitHub 按标题搜(比搜裸 ID 召回高)。多个候选时让用户确认用哪个。 (PapersWithCode 公开 API 已废弃返回 HTML,故不用;GitHub+arXiv 已覆盖常见情况。)
Stage C — 分析完整复现要求(scripts/analyze_repo.py)
python "<SKILL_ROOT>\scripts\analyze_repo.py" "<repo_url>" --json > analyze_report.json
免克隆(走 GitHub tree API + raw)读 README/依赖/训练入口/配置,grep 资源信号(n_gpu、deepspeed、 batch_size、epochs、数据集路径),产出 JSON 可行性报告:入口、依赖、GPU 需求、阻断项、建议的最小化覆盖项。
Stage D — 制定最小复现方案(agent 推理)
综合 analyze_report + 本机硬件(CPU/单卡)制定降维裁剪方案。常见手段:
- 换小模型:d_model/layer/head 全砍(如 512→64、6→2、8→2)
- 砍步数:几百 epoch → 几百步
- 关多卡:
num_processes=1、关 DeepSpeed/accelerate、必要时上 LoRA 省 VRAM - 去重型数据依赖:WMT/私有数据集 → 合成任务(序列复制/反转/排序),0 下载
- 明确要验证的核心论点:不追求复现原始 SOTA 数字,只验证论文核心主张是否成立
What ships with it
8 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- scripts/analyze_repo.py 8.5 KB runs code
- scripts/cost_estimate.py 4.6 KB runs code
- scripts/find_repo.py 4.5 KB runs code
- scripts/summarize_eval.py 4.4 KB runs code
- templates/reference/data.py 1.5 KB runs code
- templates/reference/model.py 5.5 KB runs code
- templates/reference/train.py 4.3 KB runs code
- templates/run.sh.template 754 B
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2mo ago First seen · 138 lines · 146 tokens per session scan B f7ba6ca38270
paper-repro is a skill published in the GitHub repository phanghonghao/THU-Awesome-Skills (8 stars, last pushed 3mo ago), licensed MIT. It adds 146 tokens to every session and 2,180 once invoked, about $0.0006 per session on Opus 5.5. A static security scan graded it B with 2 findings (asks for root, makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
paper-repro compared with skills from other repositories
| Mod | Type | Stars | Last push | Licence | Scan | Tokens | Works with |
|---|---|---|---|---|---|---|---|
| paper-reprophanghonghao/THU-Awesome-Skills | Skill | 8 | 3mo ago | MIT | B | 146 | Claude Code, Codex |
| instrument-data-to-allotropeanthropics/knowledge-work-plugins | Skill | 35k | yesterday | Apache-2.0 | A | 123 | Claude Code |
| matlabK-Dense-AI/scientific-agent-skills | Skill | 56k | 6d ago | MIT | A | 44 | Claude Code |
| exploratory-data-analysisK-Dense-AI/scientific-agent-skills | Skill | 56k | 6d ago | MIT | A | 83 | Claude Code |
| phylogeneticsK-Dense-AI/scientific-agent-skills | Skill | 56k | 6d ago | MIT | A | 71 | Claude Code, Codex |
| research-engineerdavila7/claude-code-templates | Skill | 33k | today | MIT | A | 43 | Claude Code, Codex |
| mapping-to-snomedmaziyarpanahi/openmed | Skill | 5.8k | today | Apache-2.0 | A | 205 | Claude Code |