LLM requests can be difficult to measure across SDKs and providers. This skill identifies live callsites, routes them through Caveman, and verifies one measured request.

拉取请求审查常常过于冗长，难以快速采取行动。此技能会将反馈格式化为简洁、可直接粘贴的评论，包含位置、问题和修复建议。

Aggregate repository profiles do not identify a safe or valuable code change by themselves. This skill tests one approved candidate against a controlled baseline.

Experiment lifecycle decisions can be unsafe when evidence or server safeguards are incomplete. This skill reviews current results and recommends one supported action without executing mutations.

Large configuration files and repeated context increase token usage. This skill reviews Caveman reports and applies measured changes with individual approval.

用户可能会忘记 Caveman 模式、触发短语和设置选项。此技能会显示一张紧凑的参考卡片，而不会更改模式或保存状态。

Finding relevant code across an unfamiliar repository can consume time and context. This skill searches broadly and returns concise, verified file and line citations.

LLM cost and quality changes are difficult to explain from aggregates alone. This skill produces bounded, trace-supported Caveman Cloud reviews without changing experiments.

Unlabeled LLM traffic hides which jobs create spend. This skill discovers repository workflows, proposes durable labels, applies approved changes, and verifies attribution.

大型项目记忆文件会在每次 AI 编码会话启动时消耗 token。此技能会压缩散文式记忆文件，同时保留技术细节并保存一份可读备份。

冗长的提交会隐藏变更原因并拖慢审查。此技能会创建简短的 Conventional Commit 消息，并仅在需要时添加上下文。

冗长回答会浪费上下文并拖慢技术工作。caveman mode 会压缩解释，同时保持代码、警告和技术含义完整。

当子代理返回冗长结果时，长时间编码会话会丢失上下文。Cavecrew 引导 Claude、Codex 和 Claude Code 使用紧凑的调查、构建和审查代理。

Encrypted repository secrets require careful commands, keys, and environment handling. This skill guides SOPS workflows through an existing pnpm secrets wrapper.

Finishing a change requires careful verification, commits, pull-request setup, and CI follow-through. This skill manages that workflow until every check passes.

Merging several open pull requests together requires careful ordering, conflict handling, and repeated verification. This skill builds one GitHub stack and merges it through repository protections.

Inconsistent engineering choices make code harder to review and maintain. This skill applies clear design, testing, error, and tooling defaults across common languages.

已完成的界面在发布前通常需要一些小的视觉调整。此技能会指导 Claude、Codex 和 Claude Code 在保留设计的同时提升清晰度。

在 Claude Code Web 会话中创建的研究内容可能分散在多个分支中。此技能可指导审查、合并、文档整理和 Issue 创建。

当 handle、edge、调整大小和滚动行为彼此脱节时，React Flow 节点可能会出问题。此技能为 Claude、Codex 和 Claude Code 提供一份聚焦的检查清单，用于稳定地修改工作流节点。