这是我在本地使用的一小批代理技能集合。每个技能都存在一个顶层目录，带有自己的SKILL.md和旁边的相关参考文件。

安装

安装方式 skills.sh：

技能

交接与接取

handoff 在会话结束时写一份交接文档给 ~/.agents/handoffs/ /。pickup 在下一次会话开始时读取最新的文档，与 git 比对，并在几行内向你简要说明。/handoff relay 还会在新的 Herdr 面板中重新启动同一代理，从而获取文档。pickup/README.md 涵盖了两者如何配合以及如何在每个线带中调用它们。handoff 起初是 Mattpocock 的 handoff 技能。

牛群流动

herd-flow 定义了通过可见的 Herdr 窗格委托给 pi 的必要协议，包括窗格生命周期、报告收割和故障处理。在启动动态工作流程前，它会参考其捆绑的路由参考，用于模型路由、报告形状和升级规则。

复制人

replicant 是一项面向外部仓库和库的以源为先的研究技能。

Agent Skills

A small collection of agent skills I use locally. Each skill lives in a top-level directory with its own SKILL.md and any supporting reference files beside it.

Install

Install with skills.sh:

npx skills@latest add oscabriel/skills

Skills

Handoff and Pickup

handoff writes a handoff doc to ~/.agents/handoffs// at the end of a session. pickup reads the newest one at the start of the next, checks it against git, and briefs you in a few lines. /handoff relay also relaunches the same agent in a new herdr pane that picks the doc up. pickup/README.md covers how the two fit and how to invoke them in each harness. handoff began as mattpocock’s handoff skill.

Herd Flow

herd-flow defines the required protocol for delegating to pi through visible herdr panes, including pane lifecycle, report harvesting, and failure handling. Before launching a dynamic workflow it consults its bundled routing reference for model routing, reporting shape, and escalation rules.

Replicant

replicant is a source-first research skill for external repositories and libraries. It teaches an agent to use durable, human-findable local clones as its primary evidence source instead of relying on stale documentation, web snippets, or generated summaries.

Replicant is useful when you ask about:

• a GitHub (or GitLab, Codeberg, etc) repository;
• an open-source library or framework;
• package internals;
• implementation details;
• architecture or design decisions in a codebase;
• examples that should be grounded in real source code.

The skill uses a local clone shelf with paths like:

~/clones///

Setup

On first use, configure the clone shelf by choosing:

• Clone root — where repositories should live, for example ~/clones.
• Update policy — whether clean existing clones can be updated automatically, whether the agent should ask first, or whether it should never update.
• Clone depth — full history or shallow clones.
• Transport — SSH or HTTPS.

After setup, Replicant records those choices in its SKILL.md config block. If a local inventory file exists, the skill checks it before searching clone paths.

Usage

Ask the agent about an external repo, library, or source-backed behavior, for example:

Use replicant to look up facebook/react. How does Suspense handle thrown promises?

Replicant should then:

1. search the local clone shelf first;
2. clone the repository if it is missing;
3. update an existing clean clone according to the configured update policy;
4. inspect the source directly;
5. answer with evidence, including the commit SHA, file paths, and line ranges when practical.

Docs to Types