研究团队需要不会污染可信基线的探索性运行。此技能会规划有边界的变体扫描、对候选项排序，并将结果界定为有限证据。

探索性研究变更可能会偏离基线，并变得难以比较。此技能会指导隔离的代码适配，并在 explore_outputs 中记录具备回滚意识的证据。

研究仓库在运行前通常需要仔细设置环境和资产。此技能会创建保守的 conda 优先计划、缓存假设和设置说明。

深度学习仓库在修改前通常很难理解。此技能会映射入口点、配置、评估路径和审查风险，而不会修改代码。

研究复现经常因为代理跳过证据、运行范围过大的命令或更改协议而失败。此技能引导 Claude、Codex 和 Claude Code 以 README 优先的方式完成复现，并生成可审计的输出。

研究团队需要有纪律的方法来测试候选想法，同时避免夸大结果。此技能用于组织来源查找、想法排序、有边界的运行以及可审计输出。

CTF OSINT clues often span disconnected public sources and media artifacts. This skill organizes practical methods for authorized identity, geolocation, archive, DNS, and web research.

Apple Bazel builds usually require local Xcode and compatible Mac hardware. This skill guides remote builds, simulator installation, uploads, verification, and troubleshooting through Limrun.

Local Apple builds require macOS and Xcode infrastructure. This skill sends projects to Limrun for remote builds, simulator installation, signing, previews, and App Store uploads.

iOS UI testing often requires macOS hardware and manual simulator access. This skill controls Limrun cloud simulators through the lim CLI from any supported environment.

Text-only DeepSeek clients cannot interpret pasted images or produce image files. This skill deploys a proxy using OCR, vision models, and generation backends.

创意团队需要快速生成图像和视频，而无需手动跟踪任务。此技能会将提示发送到 LibTV、轮询进度、上传参考文件并下载结果。

Automated browsers can expose inconsistent fingerprints that weaken test reliability. This skill measures those inconsistencies with liarjs and explains the failing checks.

Fingerprint scores often hide which component caused each failure. This skill maps liarjs checks to launch, page, network, or machine sources.

Browser fingerprint changes can silently break automated Chromium workflows. This skill adds repeatable liarjs scans, baseline comparisons, and score gates to CI pipelines.

Contradictory browser and network signals can reveal automation or spoofing. This skill runs liarjs checks and explains the resulting consistency score.

Substantial agent tasks often end with incomplete work or unsupported completion claims. Unlazy creates acceptance gates, runs checks, and records evidence before reporting completion.

通用 AI 界面通常看起来不一致，也难以细化。此技能会将项目方向转化为结构化的 DESIGN.md，用于指导 Google Stitch、Claude、Codex 和 Claude Code。

现有应用常常显得通用、不一致或未完成。此技能会审计界面，并在不改变核心技术栈的情况下应用有针对性的视觉升级。

通用 AI 生成的界面通常显得拥挤、光亮或不一致。此技能为 Claude、Codex 和 Claude Code 提供精确的视觉规则，用于打造平静的编辑风布局。