让Claude Code主动提出自己的强力工具——工作流程、目标、循环、钩子——的技能
概览
这些技能让Claude Code在自身能力上更具主动性。 Claude Code 配备了强大的力量倍增器——多代理工作流程、会话目标、循环循环、钩子、自定义技能——但你必须知道去请求它们。这些技能会颠覆一切：，只要有具体的草图和诚实的成本，你就能决定。 用OpenAI Codex？同一个侦察员，重新定位Codex的原生工具集（子代理扇动、目标、自动化、生命周期钩子），现在住在codex-power-tools。 机会发现者。当你处理普通任务时，它会观察八个形状，并在右接缝处呈现一个提案： 核心礼仪已成：每次机会一次提案，绝不堆叠提案，绝不把热情当成认可，拒绝时毫不客气地放弃。 收益形态：“我们如何系统地审计这25个端点？
概述（AI 翻译）
这些技能让Claude Code在自身能力上更具主动性。
Claude Code 配备了强大的力量倍增器——多智能体工作流程、会话目标、循环循环、钩子、自定义技能——但你必须知道去请求它们。这些技能会颠覆这一点：Claude 注意到你的任务适合它的某个电动工具并提出它，并提出它，给出一个具体的草图和诚实的费用，然后你来决定。
用OpenAI Codex？同一个侦察员，重新定位Codex的原生工具集（子代理扇动、目标、自动化、生命周期钩子），现在住在codex-power-tools。
技能
suggest-power-tools
机会发现者。当你处理普通任务时，它会观察八个形状，并在右接缝处呈现一个提案：
| 它注意到了… | 它暗示了… |
|---|---|
| 同样的问题在许多文件/项目、审计、值得对抗验证的声明中出现 | 多代理 工作流程（带有流水线草图 + 成本，绝不自动启动） |
| 几个可分离、长期存在的轨道，必须协商不断演变的合同，每个轨道都并行引导 | 一个代理团队（一个主导 + 同伴团队成员;写作者工作树隔离），绝不用于对共享文件进行紧圈调优 |
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
Claude Power Tools
Skills that make Claude Code proactively smarter about its own capabilities.
Claude Code ships with serious force-multipliers - multi-agent workflows, session goals, recurring loops, hooks, custom skills - but you have to know to ask for them. These skills flip that: Claude notices when your task fits one of its power tools and proposes it, with a concrete sketch and an honest cost, and you decide.
Using OpenAI Codex instead? The same scout, retargeted at Codex’s native tool set (subagent fan-out, goals, automations, lifecycle hooks), lives at codex-power-tools.
Skills
suggest-power-tools
The opportunity spotter. While you work on ordinary tasks, it watches for eight shapes and surfaces a proposal at the right seam:
| It notices… | It suggests… |
|---|---|
| The same question across many files/items, audits, claims worth adversarially verifying | A multi-agent Workflow (with pipeline sketch + cost, never auto-launched) |
| Several separable, long-lived tracks that must negotiate an evolving contract, each steered in parallel | An agent team (a lead + peer teammates; writers worktree-isolated), never for tight-loop tuning on shared files |
| A multi-turn mission with a mushy finish line | A /goal with an evaluator-verifiable condition, drafted ready to fire |
| Babysitting work: waiting on CI, recurring checks, “remind me later” | /loop or /schedule |
| You just did something the hard way that will recur | Capturing it as a skill while the context is hot |
| “From now on, when X…” or a repeated manual ritual | A settings hook (the one automation memory can’t do) |
| A durable design principle crystallizing mid-conversation | A one-line decision-log entry |
| Task gravity not matching the effort applied | Effort recalibration - acts on its own levers (narrating them), proposes yours (/fast, /model, budgets) |
Core etiquette baked in: one proposal per opportunity, never stacks proposals, never treats enthusiasm as approval, drops it without ceremony if you decline.
The payoff shape: a “how do we systematically audit these 25 endpoints?” aside becomes a multi-agent workflow that produces a rerunnable consistency checker and a grounded drift report - a durable tool you’d never have thought to ask for.
Install
Copy any skill directory into your skills folder:
git clone https://github.com/Matt-Hulme/claude-power-tools.git
cp -r claude-power-tools/suggest-power-tools ~/.claude/skills/
Skills follow the standard SKILL.md convention (YAML frontmatter + progressive-disclosure references), so they also work with any agent runtime that reads Claude-style skill directories. If you rename the directory, update the name: field in the frontmatter to match.
Optional: the firing hook
推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add matt-hulme/claude-power-tools