auto-production

大改动 (>3 文件) 之前先画依赖图 + 改动计划，确认后再动手。**所有外部 ref（line# / API / partition / env var）必须配 _verify 守门函数 runtime fail-loud**，不能盲信 WebFetch 结果。Use when user says "arch plan", "改动计划", "画依赖图", "重构前", "big change", "多文件改", "先规划", "monkey patch", "插入点".

Driver Claude（通常通过 /remote-drive 控制 sub-agent CC）跑实战发现新 bug / 配置问题 / 架构选择需重新拍板时，把 findings 标准化写到 .driver_findings_<round>_<context>.md，让 sub-agent 读了去 patch。比 send-keys 长 message 可靠 100 倍。Use when user says "driver findings", "findings file", "round N 反馈", "远程发现", "hpcc 验证发现", "driver patch", "sub-agent 修 bug".

用一个 Claude (driver) 通过 tmux 远程控制另一个 Claude Code session (sub-agent CC) 走完 pipeline。把 sub-agent 当 worker 用：driver brief 任务 → send-keys 输入 → capture-pane 监听 → 处理 AskUserQuestion → 等下次 idle。Use when user says "remote drive", "驱动另一个 cc", "sub-agent cc", "控制别的 claude", "tmux drive", "我已经开了 cc 你来控制".

起任何实验/sprint 前必须填 Karpathy 5-tuple: Goal / Scope / Metric / Verify / Guard。不填拒绝起训练。Stage 4 → 5 之间 RESOURCE gate 的一部分。用户说 'contract' / '5-tuple' / '签合同' / 'sprint plan' 时调用。

把本地 project 推到远程 compute host（hpcc / bcc / tasl-labserver）+ 远端 bootstrap auto-production framework + 验证 stubs import 通。给 driver 用，跟 /cross-host-sync (run sync) 互补——后者管 run state，本 skill 管 code state。Use when user says "sync to remote", "部署到 hpcc", "推到集群", "rsync project", "远程跑", "hpcc 上验证", "deploy 上去".

写完代码先用 AST parser 做静态分析（import 完整性、签名匹配、未定义符号），过了才执行。防 "跑了 30 秒才发现 ImportError"。Use when user says "ast validate", "静态分析", "先检查", "语法检查", "import 验证", "static check".

标准化失败 retrain 循环：scancel → audit (sub-agent) → patch → 重起。这次 hf-jax 做了 v3/v4/v5/v6 4 次的模板化。Use when user says "audit retrain", "重训", "训挂了", "scancel audit patch", "diagnose retrain", "为什么 SR 低".

决策为 REFINE 或 PIVOT 时，自动 snapshot stage-N_v{K}/ 目录（保留旧版本），新 run 用 3-part prompt: CONTINUATION (上次进度) + PRIOR PLAN (上次方案) + DELTA (这次改动)。Use when user says "auto version", "snapshot", "备份旧版本", "v3 v4 v5", "version run", "保留旧 run".

Benchmark 选择自动化：Surveyor (调研) → Selector (筛选) → Acquirer (下载) → Validator (验证完整性) 4 阶段 sub-agent pipeline。Use when user says "benchmark agent", "选 benchmark", "benchmark 调研", "surveyor selector", "找 dataset", "benchmark pipeline".

Agent 收到 user proposal 不要张口就 "好的，收到，马上做"。先评估：(a) user 是否充分理解 trade-off (b) 有没有更优方案 (c) 是否违反前面设定的 Guard。然后才同意 / 反对 / 提替代。反 sycophancy。Use when user says "concession threshold", "不要 yes-man", "先想想再答", "push back", "质疑我一下".

每个下载的 ckpt / dataset / paper 记录 obtained_via (URL/script) + obtained_at (timestamp) + sha256 + license + 来源 paper。重现 run 时知道每个 dependency 精确版本。Use when user says "corpus schema", "登记下载", "literature corpus", "sha256", "download manifest", "reproducibility".

针对机器人 / VLA 领域建一个 datalake：数据集元数据 + 常用 API + 行话词典 + 已知坑。Biomni 模板。下次同领域 project 直接复用。Use when user says "datalake", "know-how", "领域词典", "常用 dataset 库", "domain library", "avoid known pitfalls".

重大决策时让 N 个 agent 用 N 种立场辩论，再用 blind judge 给结论。比 cross-review 更结构化（明确的 pro/con/judge 三方）。Use when user says "debate", "辩论", "几个 agent 各执一词", "blind judge", "pro con", "重大决策".

每个 project 一个 Dockerfile + GPU pinning via .env，环境彻底隔离。避免 conda env 互相污染。Use when user says "docker", "docker workplace", "容器化", "env 隔离", "GPU pin", "Dockerfile".

Agent 主动输出 confidence rating 0-5 + 列出可疑点。把 user 频繁说的 "double check 一下" 固化。Use when user says "double check", "再确认", "你确定吗", "confidence", "自查", "reliability".

AI Research 7 大 known failure mode checklist：data snoop / distribution mismatch / leakage / weak baseline / single-seed / pre-trained sanity / implementation diff。Stage 推进前必须逐项 ✓。Use when user says "failure checklist", "7 种坑", "data snoop", "leakage check", "sanity check", "before submission check".

所有 findings 一个 dict + 一个可视化 graph（谁影响谁，主线 vs 支线）。替代 findings.md 纯线性 append。Use when user says "findings map", "研究地图", "finding graph", "主线支线", "insight network", "知识图".

Meta-gate-check skill：给定 gate 名 (NOVELTY / METHOD / RESOURCE / RESULTS / FINAL)，自动跑 checklist 检查，全 ✓ 才放行。代替 ad-hoc agent 自己判 "我觉得过了"。Use when user says "gate check", "过 gate", "check NOVELTY", "check METHOD", "check RESULTS", "过 FINAL", "can I proceed".

启动 project 时选 1 个 HITL 模式全程跑：full-auto / gate-only / checkpoint / co-pilot / step-by-step / custom。整个 project 按这个 mode 跑，不混乱。Use when user says "hitl mode", "co-pilot mode", "full auto", "checkpoint mode", "切模式", "interactivity".

STORM-style Perspective-Guided Question Asking. 找 5-10 篇 similar-topic paper → 提取每篇独特视角 → 用这些视角驱动 question 自动生成。破单一视角盲点。Use when user says "找视角", "perspective", "STORM idea", "广角调研", "找问题方向", "idea brainstorm", "换角度想".

Simulated Conversation (writer × expert)。Agent 扮演 writer 和 expert 两个角色互相 role-play 深挖一个 topic。比单一 monologue 出来的 idea 更扎实。Use when user says "sim conversation", "对话深挖", "扮演专家", "两个 agent 聊", "role-play idea", "思辨".

为 reviewer 找茬准备反驳武器库。给一个 claim，agent 列出 5 种最可能的攻击角度 + 每个怎么反驳，附 evidence 链接。Use when user says "kill argument", "反驳武器", "reviewer 会怎么挑", "adversarial defense", "theory paper 防御".

Agent 发现 lesson 时打 [LEARN:notation] / [LEARN:code] / [LEARN:method] 等 tag，自动入 MEMORY.md 对应 tier。跨 session 自动可被检索。Use when user says "learn tag", "记一下", "memory tag", "lesson learned", "保存到 memory", "长期记忆".

Idea 拍下来后自动去 Semantic Scholar / OpenAlex 搜，看是否已有 paper 做过。防 "想了 3 周才发现 ICLR'25 有人做过"。Use when user says "novelty check", "查重", "有人做过吗", "prior art", "是否新", "是不是 novel", "重复了没".

10-mode academic-paper writing skill：intro / related-work / method / experiment / discussion / abstract / conclusion / appendix / reviewer-defense / one-pager。一个 skill 多模式切换。Use when user says "paper mode", "写 intro", "写 method", "写 abstract", "写 related work", "写 conclusion", "academic paper section".

Paper 写作 6-stage 流水线：outline → draft → review → revise → format → cite-verify。每 stage 出 deliverable，不一气呵成。Use when user says "paper pipeline", "写 paper", "投稿流水", "6-stage paper", "写论文", "start paper".

Paper 接收后生成 3'×4' A0 海报 PDF + editable PPTX + SVG。venue-specific 布局模板。Use when user says "paper poster", "做 poster", "海报", "A0 poster", "会议海报".

Paper 接收后生成 10 页 slides（Beamer + PPTX），含 speaker notes。可单独跑（不跟 /paper-talk 绑定）。Use when user says "paper slides", "做 slides", "PPT", "Beamer slides", "15 页 ppt", "slides only".

Paper 接收后一行命令生成 12-min talk script + Beamer slides + speaker notes + Q&A prep。Use when user says "paper talk", "做 talk", "会议 talk", "演讲稿", "presentation", "我要做 oral".

Start project 时用 8 种 stakeholder 视角（reviewer / PI / 同领域学生 / 跨领域学者 / 工业界 / 政策制定者 / 中文用户 / 复现者）各提 3 个问题，作为后续 AskQuestion 的弹药库。Use when user says "persona probe", "8 stakeholder", "reviewer 视角", "各角度问题", "stakeholder questions", "想问题".

N 个连续 sprint metric 不动（如 SR 卡在 60% 不上不下），自动弹窗建议 PIVOT 而不是继续 REFINE。防 sunk cost 心理一直 "再试试"。Use when user says "plateau", "卡住了", "metric 不涨", "sunk cost", "该不该继续", "是不是该 pivot".

Publication-ready figure + LaTeX↔Excel table 切换。NeurIPS / CoRL / ICLR 模板内置。把 /auto-viz 的 PNG 升级成投稿质量的 PDF/SVG。Use when user says "pubfig", "publication figure", "投稿质量图", "latex table", "美化图", "出 PDF figure".

Reviewer 评论流水线：分类 → 找 supporting evidence → 写 rebuttal → cross-model review → submit。3 道 safety gate：no fabrication / no overpromise / full coverage。Use when user says "rebuttal", "reviews", "reviewer 来了", "反驳", "答辩", "OpenReview comment".

下载 ckpt / dataset 前显式列：总大小 / 目标盘剩余空间 / 估计下载时间 / 是否需要 paywall / 是否要 import cookie。User 拍板再下。防 "下到一半盘满"。Use when user says "resource planning", "下载前", "估空间", "plan download", "准备下数据", "check disk".

Workflow 5: text-only resubmit across venues。port 已 polish paper 到新 venue (no new experiments / no bib edits / no framework changes)。5 阶段：物理隔离 → 5-layer 匿名性 → audits (--soft-only) → microedits → adversarial gate → final compile。Use when user says "resubmit", "换 venue", "投另一个会", "重新投稿", "port to ICLR", "text-only resubmit".

任何单文件超 400 行触发提醒，超 800 行强制 refactor 才能继续。防代码膨胀失控。Use when user says "simplify gate", "文件太大", "line limit", "refactor 提醒", "代码膨胀".

固定 6 角色 self-evolving team：PI / postdoc / phd-student / engineer / reviewer / writer。每周 team review 一次，相互调整 prompt。仿真一个真实科研小组。Use when user says "six agent", "6 角色", "team review", "虚拟 lab", "PI postdoc", "role play team".

sub-agent fork 模板：给 prompt 模板 + 输出格式约定 + 资源约束 + 完成检测，fork 一个 Agent sub-task 后台跑。比每次手写 Agent prompt 省 10x。Use when user says "spawn task", "fork agent", "开个 sub-agent", "并行任务", "background agent", "派活".

Phase 入口给 agent 一份 YAML: compute_budget / dataset_loader_code / eval_rules / deadline。Agent 必须读完才能动。固化资源分配。Use when user says "task notes", "phase yaml", "资源声明", "agent 前置读", "compute budget".

把所有 v1/v2/v3/v6 run 画成树状图（含 branch 关系），每个分支 metric 一目了然，hover 看 hyperparam diff。替代爬不同 wandb run 人脑对比。Use when user says "tree viz", "run 树", "对比 runs", "v1 v2 v3 v6 关系", "branch tree", "看祖宗".