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 Lambenthan/paper-discipline-skills --skill paper-pilot-before-batchgit clone --depth 1 https://github.com/Lambenthan/paper-discipline-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/lambenthan/paper-discipline-skills/paper-pilot-before-batch)<a href="https://agentmods.dev/skills/lambenthan/paper-discipline-skills/paper-pilot-before-batch"><img src="https://agentmods.dev/badge/skills/lambenthan/paper-discipline-skills/paper-pilot-before-batch/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/lambenthan/paper-discipline-skills/paper-pilot-before-batch"><img src="https://agentmods.dev/badge/skills/lambenthan/paper-discipline-skills/paper-pilot-before-batch.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 | $0.00119 | $0.01400 |
| Opus 5 | $0.00060 | $0.00700 |
| Sonnet 5 | $0.00024 | $0.00280 |
| Haiku 4.5 | $0.00012 | $0.00140 |
Grade A, and why
paper-pilot-before-batch scanned grade A with 0 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 11d 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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
paper-pilot-before-batch：批量任务先小样本
核心理念
批量任务的可怕之处：跑错 1 个 = 跑错 100 个。 小样本花 1 分钟，全量跑错回滚要 1 小时。
批量是放大器。脚本对了，放大效率；脚本错了，放大灾难。
触发条件
满足任一条 → 触发：
- 用户说"全部"、"所有"、"每一个"、"逐个"、"批量"、"统一"
- 操作对象数量 ≥ 30（文件 / 段落 / 引用 / 行）
- 操作对象数量未知（"整个目录的"、"全篇的"、"剩下的"）
- 涉及修改性操作（不是只读 / 查询）
强制流程
检测到批量操作
│
▼
随机抽 3–5 个样本（不要只取前 3 个，前 3 个常常是规整数据）
│
▼
只对样本执行
│
▼
把样本结果展示给用户：
- 每个样本的"改动前 vs 改动后"
- 每个样本的"判断依据"（为什么这样改）
│
▼
明确问用户：
「这是 N 个样本的改法，逻辑对的话我推全量（共 M 个）。
有不对的请指出，我调整后重新跑样本。」
│
▼
确认 → 全量（≥ 30 时配合 paper-parallel-audit 用并行 Agent）
不对 → 改逻辑 → 重新跑样本
抽样策略（不是随便抽）
抽样要覆盖边界情况：
- 1 个最常见的情况
- 1 个看起来可能出问题的情况（最长 / 最短 / 含特殊符号 / 跨行）
- 1 个完全不该被处理的情况（验证你的判定逻辑不会误伤）
不要 head -3 —— 前 3 个常常是规整数据，跑过样本不代表跑过全量。
标准回复模板
这是个批量操作，共 M 个对象。 先抽 5 个样本跑一下，包括 1 个常见情况、1 个边界情况、1 个不该被改的情况。
[样本 1] 改前 → 改后，依据：… [样本 2] … [样本 3] …
5 个看着对的话，我就推剩下的 M-5 个。 有看着不对的请指出来，我调整后重跑样本。
❌ 反例（书 §10）
用户：「把所有引用的格式按 GB/T 7714 统一一下。」（共 156 条）
错误做法：写一个正则脚本，一次性跑 156 条——结果发现：
- 部分引用里的"等"被当成 et al. 多处理了一次
- 部分中文姓名拼音被错拆
- 全部 156 条都错，且原始格式没存
正确做法：先跑 5 条样本（含 1 条中文姓名、1 条多作者带"等"、1 条单作者）→ 用户检查 → 修正脚本 → 再推 156 条。
"可以跳过样本" 的极少数情况
- 纯只读：批量统计、批量列出（不修改任何文件）
- 逐条用户确认：每条都让用户点头才下一条（实质就是没批量）
- 数量 < 30 且每条 < 1 秒：直接做完，结果可视
Rationalization Table
| 念头 | 现实 |
|---|---|
| "我已经在脑子里推演过了，肯定对" | 推演 ≠ 实测。书里的 156 条惨案，AI 也"推演过" |
| "用户着急，跳过样本更快" | 跑样本 1 分钟 vs 跑错回滚 1 小时 |
| "这个任务很简单，正则一行就搞定" | 简单的正则吃掉的都是边界情况 |
| "我跑前 3 个验证一下" | 前 3 个不是样本，是规整数据 |
| "我备份了，跑错了能回滚" | 回滚是兜底，不是策略 |
| "用户说他相信我直接全量跑" | 仍然先 5 个样本，跑出来给他看 |
Red Flags
- 你即将对 ≥ 30 个对象 / 文件循环调用 Edit / Write → 停，先样本
- 你写了一段 for 循环 / 正则替换准备 apply 全量 → 停，先样本
- 你想"先跑一半看看" → 一半不是样本，跑错就 50 个错的
- 用户说"直接跑全量" → 仍然先 5 个样本，把结果给他看
配套 skill
- 全量阶段如果 ≥ 30，配合
paper-parallel-audit用并行 Agent - 修改性操作前必须
paper-confirm-before-doing已经确认过方案
来源
《Claude Code 科研手记》第 10 章「并行 Agent」、§4.2「批量文献核查」
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.
- 11d ago First seen · 140 lines · 119 tokens per session scan A 6dd3d54bdaa3
paper-pilot-before-batch is a skill published in the GitHub repository Lambenthan/paper-discipline-skills (19 stars, last pushed 3mo ago), licensed MIT. It adds 119 tokens to every session and 1,400 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
secure-auth
Secure authentication patterns (OWASP, NIST). Use for login, registration, password reset, sessions, JWT, OAuth, MFA, passkeys.
academic-writing
Scholarly writing and research compliance. Use for CRediT, preregistration, Plan S, Nelson Memo, preprints, ORCID, LLM disclosure.
page-monitoring
Web page change detection, availability tracking, and RSS feed generation. Use to monitor changes, downtime, or make a feed.
web-archiving
Web archiving and retrieval via Wayback Machine and Archive.today. Use to preserve content, reach dead pages, or save evidence.
accessibility-compliance
Web accessibility patterns for news and academic sites. Use for WCAG audits, alt text, accessible data viz, and assistive tech.
one-way-door
Flags irreversible decisions before commit. Use for data models, infra, auth boundaries, API contracts, event schemas, CI/CD.