# 17 Skills I Would Install on a Fresh Hermes Setup (10x Powerful Hack)
**作者**: Farea
**日期**: 2026-08-23T14:49:48.000Z
**来源**: [https://x.com/FareaNFts/status/2091538403611533820](https://x.com/FareaNFts/status/2091538403611533820)
---
If I wiped Hermes tomorrow and started from zero, I would not open a blank chat and "figure it out."
I would install skills first.
A raw Hermes install is smart. A skilled Hermes setup is unfair. Skills are the difference between a chatbot that answers questions and an agent that researches, codes, edits video, cleans writing, runs loops, and keeps ten jobs moving without you living in the terminal.
This is the exact set I would load. Not every skill on the internet. The 17 that cover almost every real job I give an agent.
What each one does. When to install it. What to skip on day one.
| # | Tool | One-line job | Stars (approx) |
| --- | --- | --- | --- |
| 1 | Agent-Reach | Give your agent internet eyes without paid APIs | ~73k |
| 2 | make-interfaces-feel-better | Fix the tiny UI details that make apps feel cheap | ~3k |
| 3 | oh-my-hermes | Multi-agent planning + verified execution for Hermes | ~270 |
| 4 | Anthropic Cybersecurity Skills | Huge security skill library for agents | ~30k |
| 5 | OpenMontage | Turn your coding agent into a video production crew | ~49k |
| 6 | Minions | Kanban mission control for Hermes tasks | ~625 |
| 7 | Resemble Detect | Spot AI fake audio/image/video | small / niche |
| 8 | addyosmani/agent-skills | Spec → plan → build → test → ship engineering skills | ~89k |
| 9 | Composio skills | Connect agents to real apps (auth, tools, triggers) | Composio ecosystem |
| 10 | youtube-full | Transcripts, search, channels without yt-dlp pain | ~540 |
| 11 | Humanizer | Strip AI writing tells from drafts | popular skill |
| 12 | Defuddle | Clean web pages into readable Markdown | ~9k |
| 13 | Matt Pocock skills | Small composable skills for real engineering | ~224k |
| 14 | SkillClaw | Skills that improve themselves after real work | ~2.5k |
| 15 | Browser Harness | Agent drives your real browser and learns helpers | ~17k |
| 16 | codebase-memory-mcp | Index your repo as a knowledge graph for agents | ~40k |
| 17 | Loopy / Loop Library | Repeatable agent loops with checks and stop rules | ~3k |
## What a “skill” even is
Think of a skill like a recipe card your agent can pull when needed.
Without skills:
- you re-explain the same workflow every session
- quality depends on the model’s mood
- multi-step work drifts
With skills:
- the steps are written down once
- the agent loads them when the task matches
- you can share the same recipe across projects
Hermes already ships with many skills. These 17 are the external packs people bolt on when they want a serious setup.
## 1. Agent-Reach
Repo: [Panniantong/Agent-Reach](https://github.com/Panniantong/Agent-Reach) Stars: ~73k
What it is
A CLI that gives your agent eyes on the public internet - X/Twitter, Reddit, YouTube, GitHub, and more - without you wiring a paid API for every platform.
Problem it fixes
Your agent can write code. Ask it to “check Twitter reactions” or ''summarize this YouTube tutorial'' and it often fails: blocked scrapers, paid APIs, login walls, garbage HTML.
How it helps newbies
Install once. Then say things like:
- search X for what people say about this tool
- pull the YouTube transcript and summarize it
- check Reddit for the same bug
When to use it
Content research, competitor checks, social listening, tech support research.
## 2. make-interfaces-feel-better
Repo: [jakubkrehel/make-interfaces-feel-better](https://github.com/jakubkrehel/make-interfaces-feel-better) Stars: ~3k
What it is
An agent skill focused on UI polish: animation, type, icons, hover states, hit areas, shadows, optical alignment - the stuff that makes an interface feel intentional.
Problem it fixes
AI-built UIs often “work” but feel off. Wrong spacing. Dead hover states. Buttons too small. Text that looks auto-generated.
How it helps newbies
After your agent builds a page, load this skill and ask it to pass over the UI for feel, not features.
When to use it
Landing pages, dashboards, internal tools, any frontend the agent just shipped.
Install idea
```
npx skills add jakubkrehel/make-interfaces-feel-better
```
## 3. oh-my-hermes (OMH)
Repo: [witt3rd/oh-my-hermes](https://github.com/witt3rd/oh-my-hermes) Stars: ~270
What it is
Multi-agent orchestration skills built for Hermes, inspired by oh-my-claudecode. Skills for research, interviews, consensus planning, and verified execution.
Core skills inside
| Skill | Job |
| --- | --- |
| omh-deep-research | Split research → parallel search → synthesize → check citations |
| omh-deep-interview | Socratic requirements interview |
| omh-ralplan | Planner / architect / critic debate until a plan holds |
| omh-ralph | Implement → verify → loop until done |
| omh-autopilot | Full pipeline of the above |
Problem it fixes
One agent freestyling a big task often skips requirements, weak-plans, then “finishes” without proof.
When to use it
Unfamiliar domains, multi-day features, anything where you want plan quality before code.
Suggested flow
```
research → interview → ralplan → ralph
```
## 4. Anthropic Cybersecurity Skills
Repo: [mukul975/Anthropic-Cybersecurity-Skills](https://github.com/mukul975/Anthropic-Cybersecurity-Skills) Stars: ~30k
What it is
A huge open library of structured cybersecurity skills for agents (~817 skills, many security domains). Mapped to frameworks like MITRE ATT&CK, NIST CSF, ATLAS, D3FEND, AI RMF, and fraud frameworks.
Important honesty
This is a community project, not an official Anthropic product. It includes offensive/dual-use techniques meant for authorized testing, defense, research, and education only.
Problem it fixes
Your coding agent is not a senior security analyst by default. These skills give it playbooks for investigations, detection, forensics-style workflows, and framework-aligned guidance.
When to use it
Security reviews, incident response drafts, threat modeling, authorized red-team prep, learning security workflows with an agent tutor.
Newbie rule
Only point it at systems you own or have written permission to test.
## 5. OpenMontage
Repo: [calesthio/OpenMontage](https://github.com/calesthio/OpenMontage) Stars: ~49k
What it is
An open-source agentic video production system. Pipelines, tools, and production-knowledge files that turn a coding agent into something closer to a small video crew.
Problem it fixes
Video work is a pile of tools: script, b-roll, edit, captions, export. Most agents can “talk about video.” OpenMontage aims at actual production workflows.
When to use it
Short-form content, explainers, product demos, YouTube-style production with an agent in the loop.
Who it’s for
Creators and marketers who already use agents daily and want media output, not just text.
## 6. Mission Control / Minions
Repo: [agent37-platform/minions](https://github.com/agent37-platform/minions) Stars: ~625
What it is
A Kanban-style control panel for Hermes work. Create tasks, watch them run, review results, keep long jobs from disappearing into terminal history.
Problem it fixes
Hermes is strong. Managing five parallel sessions in raw terminals is not. You forget what finished, what’s stuck, and what needs your approval.
Features that matter
- board: in progress / review / done
- live streaming of tool calls
- human sign-off before “done”
- per-task model settings
- scheduled Hermes jobs
- local-first option (SQLite, data on your machine)
Quick start
```
npx minionsai
```
Then open http://localhost:6969.
When to use it
Anyone running Hermes as a real worker, not a one-off chat toy.
## 7. Resemble AI Detect
Repo: [resemble-ai/detect-skill](https://github.com/resemble-ai/detect-skill)
What it is
An agent skill for media safety - detecting AI-generated or manipulated audio, images, and video using Resemble’s detection stack.
Problem it fixes
Deepfakes and synthetic media are everywhere. If your agent handles UGC, brand safety, or research, you need a check step before you trust a clip.
When to use it
Content moderation pipelines, brand safety reviews, “is this clip real?” checks before you quote or repost something.
Honest note
Treat detection scores as signals, not courtroom proof. Always keep a human judgment step for high-stakes calls.
## 8. addyosmani/agent-skills
Repo: [addyosmani/agent-skills](https://github.com/addyosmani/agent-skills) Stars: ~89k
What it is
Production engineering skills from Addy Osmani’s world: a lifecycle of define → plan → build → verify → review → ship.
Commands mental model
| Command idea | Meaning |
| --- | --- |
| /spec | Define what to build |
| /plan | Break into small tasks |
| /build | Implement in slices |
| /test | Prove it works |
| /review | Improve code health |
| /webperf | Measure before optimize |
| /ship | Get it live carefully |
Problem it fixes
Vibe coding without gates. These skills force the habits senior engineers already use: specs, small tasks, tests, review.
When to use it
Real products, not throwaway scripts. Especially frontend/web work and multi-day features.
Install idea
```
npx skills add addyosmani/agent-skills
```
## 9. Composio skills
Repo: [ComposioHQ/skills](https://github.com/ComposioHQ/skills)
What it is
Distributable skills that teach agents how to build with Composio - Tool Router, auth, sessions, triggers, production patterns.
Problem it fixes
Connecting an agent to Gmail, Slack, GitHub, CRMs, etc. is messy. Composio is the integration layer; these skills teach the agent the safe patterns (user IDs, sessions, webhooks, framework wiring).
When to use it
Any Hermes/agent setup that needs real app actions, not just local files.
Install idea
```
npx skills add composiohq/skills
```
## 10. youtube-full (YouTube skills)
Repo: [ZeroPointRepo/youtube-skills](https://github.com/ZeroPointRepo/youtube-skills) Stars: ~540
What it is
Skills so your agent can:
- pull YouTube transcripts
- search videos
- browse channels
- extract playlists
Without fighting yt-dlp blocks on cloud IPs.
Problem it fixes
“Summarize this YouTube video” is one of the most common agent fails. This package is built for that job.
Hermes install idea
```
hermes skills install skills-sh/ZeroPointRepo/youtube-skills/skills/youtube-full
```
When to use it
Learning from tutorials, content research, competitor video breakdowns, course notes.
Free tier note
The underlying transcript API advertises free credits on signup. Confirm current limits on their site.
## 11. Humanizer
Repo: [blader/humanizer](https://github.com/blader/humanizer)
What it is
A skill that rewrites AI-sounding text so it reads like a person wrote it. Based on Wikipedia’s “Signs of AI writing” patterns.
Problem it fixes
Your agent drafts are useful and still sound like a press release: “pivotal,” “landscape,” “it’s not X, it’s Y,” perfect three-item lists, empty hype.
How it works in practice
1. first rewrite without protecting the AI structure
2. check remaining AI tells
3. rewrite again
4. keep facts; don’t invent names/numbers
When to use it
Every public post, docs page, client email, or launch note the agent wrote.
Pro tip
Paste 2-3 paragraphs of your own writing and ask it to match your voice.
## 12. Defuddle
Repo: [kepano/defuddle](https://github.com/kepano/defuddle) Stars: ~9k
What it is
A library/CLI that takes a messy web page and returns the main content as clean HTML or Markdown. Built by the Obsidian Web Clipper world (kepano).
Problem it fixes
Agents fetch pages and choke on nav bars, ads, cookie banners, and nested HTML. Defuddle strips the junk.
When to use it
Research, clipping articles into notes, feeding clean context into Hermes, building your own reader pipeline.
Simple mental model
URL or HTML in → readable article out.
## 13. Matt Pocock skills
Repo: [mattpocock/skills](https://github.com/mattpocock/skills) Stars: ~224k
What it is
Matt’s daily engineering skills: small, composable, editable. Built for real apps, not “vibe coding theater.”
Philosophy in one line
Big process frameworks take control away from you. These skills stay small so you can hack them.
Problem it fixes
Either your agent has no process, or it has a giant process you can’t debug. Matt’s set sits in the middle.
When to use it
TypeScript/app engineering, ticket triage, planning, and day-to-day shipping with an agent.
Install idea
```
npx skills@latest add mattpocock/skills
```
Then run the setup skill once per repo.
## 14. SkillClaw
Repo: [AMAP-ML/SkillClaw](https://github.com/AMAP-ML/SkillClaw) Stars: ~2.5k
What it is
A system where skills evolve from real use. You chat normally. In the background it digests sessions, improves skills, dedupes junk, and compounds experience across agents/devices.
Problem it fixes
Your Hermes skill folder becomes a junk drawer: duplicates, half-baked notes, stale steps. Nobody “digests” the learning.
When to use it
If you live in Hermes daily and want the skill library to get sharper over time without manual gardening every week.
Honest framing
This is infrastructure for power users. Worth it after you already have a base skill set.
## 15. Browser Harness (honourable mention)
Repo: [browser-use/browser-harness](https://github.com/browser-use/browser-harness) Stars: ~17k
What it is
A self-healing browser harness. The LLM connects to your real browser over CDP, does the task, and writes missing helpers when a step isn’t supported yet.
Problem it fixes
Static browser tools break the moment a site changes. This approach improves the harness while working.
Example task energy
“Open my X profile, grab my latest 20 video posts, download them.”
When to use it
Logged-in browser work, repetitive web chores, anything that needs a real session cookie stack.
Caution
Giving an agent your real browser is powerful and risky. Start with a separate profile for experiments.
## 16. codebase-memory-mcp
Repo: [DeusData/codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp) Stars: ~40k
What it is
A high-performance MCP server that indexes your codebase into a persistent knowledge graph. Agents query structure instead of rereading 40 files every session.
Why that matters
File-by-file exploration burns tokens and still misses architecture. Graph queries answer:
- where is this used?
- what calls what?
- what breaks if I change this?
Claims to know
- 158 languages
- very fast indexing
- big token savings vs naive exploration
- local processing (code stays on your machine)
When to use it
Medium/large repos, multi-service codebases, any Hermes coding work where context thrash is killing you.
## 17. Loop Library / Loopy (honourable mention)
Repo: [Forward-Future/loopy](https://github.com/Forward-Future/loopy) Stars: ~3k
What it is
Two parts:
1. a public catalog of agent loops
2. Loopy, a skill that helps find, audit, adapt, craft, and run those loops
What a “loop” is
A one-shot prompt says: do the thing once. A loop says: do a step, measure, keep only if it helped, repeat until a stop condition.
Example:
> Find the slowest page, improve one thing, measure again. Keep the change only if it helps. Stop when targets are met or gains die.
When to use it
Perf work, flaky tests, docs cleanup, recurring ops - anything where the first attempt is never the final answer.
Key idea
Loops are powerful because they are bounded. Clear check. Clear stop. Human handoff when judgment is needed.
## How these 17 fit together (simple stacks)
Starter Hermes stack
1. Agent-Reach - internet access
2. youtube-full - video learning
3. Defuddle - clean page text
4. Humanizer - clean public writing
5. Minions - see all jobs on a board
Serious coding stack
1. codebase-memory-mcp - stop rereading the repo
2. addyosmani/agent-skills or Matt Pocock skills - engineering process
3. oh-my-hermes - plan/verify multi-agent flow
4. make-interfaces-feel-better - UI pass at the end
Content / media stack
1. OpenMontage - video production
2. youtube-full - source material
3. Humanizer - captions and posts
4. Resemble Detect - media authenticity checks
Power-user growth stack
1. SkillClaw - skills improve from real use
2. Loopy - turn repeated work into loops
3. Browser Harness - real browser chores
4. Composio skills - app integrations
## Install order if you’re new
Don’t install all 17 on day one.
Day 1
- Agent-Reach
- Humanizer
- youtube-full
Day 2
- Minions (if you run multiple Hermes jobs)
- Defuddle
- one engineering pack (Matt or Addy)
Day 3+
- codebase-memory-mcp for real repos
- oh-my-hermes when tasks get big
- OpenMontage / Browser Harness only when you need those jobs
More skills is not automatically better. A small set you actually trigger beats a junk drawer.
## Security and common sense
A few of these touch real browsers, real codebases, or security tradecraft.
Simple rules:
- only run offensive security skills on systems you’re allowed to test
- separate browser profiles for autonomous browser agents
- read install scripts before piping them into a shell
- keep secrets out of prompts and skill logs
- prefer local-first tools when the data is sensitive
A fresh Hermes agent with this set is a smart system:
- internet research without babysitting APIs
- repo memory that does not thrash tokens
- engineering process that does not vibe-code your app into a hole
- content tools that can watch video and clean drafts
- a board so ten jobs do not disappear
- loops and evolving skills so the setup gets sharper over time
If I only had one hour on a new machine, I would install in this order:
1. Agent-Reach
2. Humanizer
3. youtube-full
4. Matt Pocock skills or addyosmani/agent-skills
5. codebase-memory-mcp
6. Minions
Then I would ship real work for a week before touching the rest.
Do not collect skills like stickers. Install the ones that remove the pain you feel this week.
Blind online → Agent-Reach Fake drafts → Humanizer Session chaos → Minions Repo thrash → codebase-memory-mcp Weak engineering process → Matt or Addy Video work → OpenMontage Real browser chores → Browser Harness
That is the fresh setup I would trust
if you found this helpful give a follow. i post free apis and agentic workflows every week.
## follow [@fareanfts](https://x.com/@fareanfts) for more 🙌
## 相关链接
- [Farea](https://x.com/FareaNFts)
- [@FareaNFts](https://x.com/FareaNFts)
- [25K](https://x.com/FareaNFts/status/2091538403611533820/analytics)
- [Panniantong/Agent-Reach](https://github.com/Panniantong/Agent-Reach)
- [jakubkrehel/make-interfaces-feel-better](https://github.com/jakubkrehel/make-interfaces-feel-better)
- [witt3rd/oh-my-hermes](https://github.com/witt3rd/oh-my-hermes)
- [mukul975/Anthropic-Cybersecurity-Skills](https://github.com/mukul975/Anthropic-Cybersecurity-Skills)
- [calesthio/OpenMontage](https://github.com/calesthio/OpenMontage)
- [agent37-platform/minions](https://github.com/agent37-platform/minions)
- [resemble-ai/detect-skill](https://github.com/resemble-ai/detect-skill)
- [addyosmani/agent-skills](https://github.com/addyosmani/agent-skills)
- [ComposioHQ/skills](https://github.com/ComposioHQ/skills)
- [ZeroPointRepo/youtube-skills](https://github.com/ZeroPointRepo/youtube-skills)
- [blader/humanizer](https://github.com/blader/humanizer)
- [kepano/defuddle](https://github.com/kepano/defuddle)
- [mattpocock/skills](https://github.com/mattpocock/skills)
- [AMAP-ML/SkillClaw](https://github.com/AMAP-ML/SkillClaw)
- [browser-use/browser-harness](https://github.com/browser-use/browser-harness)
- [DeusData/codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp)
- [Forward-Future/loopy](https://github.com/Forward-Future/loopy)
- [@fareanfts](https://x.com/@fareanfts)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [10:49 PM · Aug 23, 2026](https://x.com/FareaNFts/status/2091538403611533820)
- [25K Views](https://x.com/FareaNFts/status/2091538403611533820/analytics)
---
*导出时间: 2026/8/24 14:30:22*
---
## 中文翻译
# 我会在全新 Hermes 安装上部署的 17 个技能（10 倍强大黑科技）
**作者**： Farea
**日期**： 2026-08-23T14:49:48.000Z
**来源**： [https://x.com/FareaNFts/status/2091538403611533820](https://x.com/FareaNFts/status/2091538403611533820)
---
如果我明天把 Hermes 清空、从零开始，我不会打开一个空白聊天窗口然后“慢慢摸索”。
我会先安装技能。
一个裸装的 Hermes 很聪明。一个配置了技能的 Hermes 则强得离谱。技能就是“回答问题的聊天机器人”与“能够调研、写代码、剪视频、润色文字、执行循环、同时推进十项任务而不用你一直守在终端前”的智能体之间的区别。
这正是我会安装的那一套。不是互联网上的所有技能，而是能覆盖我交给智能体的几乎每一项实际工作的 17 个。
每个技能是干什么的、什么时候安装、第一天该跳过什么，全都在这里。
| # | 工具 | 一句话用途 | 星标数（约） |
| --- | --- | --- | --- |
| 1 | Agent-Reach | 让你的智能体无需付费 API 就能“看到”互联网 | ~73k |
| 2 | make-interfaces-feel-better | 修复让应用显得廉价的小 UI 细节 | ~3k |
| 3 | oh-my-hermes | 为 Hermes 提供多智能体规划 + 已验证的执行 | ~270 |
| 4 | Anthropic Cybersecurity Skills | 面向智能体的大型安全技能库 | ~30k |
| 5 | OpenMontage | 把你的编程智能体变成视频制作团队 | ~49k |
| 6 | Minions | Hermes 任务的看板式任务控制台 | ~625 |
| 7 | Resemble Detect | 识别 AI 伪造的音频/图像/视频 | 小众 |
| 8 | addyosmani/agent-skills | 规格 → 计划 → 构建 → 测试 → 发布的工程技能 | ~89k |
| 9 | Composio skills | 将智能体连接到真实应用（认证、工具、触发器） | Composio 生态 |
| 10 | youtube-full | 无痛获取字幕、搜索、频道，不再折腾 yt-dlp | ~540 |
| 11 | Humanizer | 去除草稿中的 AI 写作痕迹 | 热门技能 |
| 12 | Defuddle | 将网页清洗为可读的 Markdown | ~9k |
| 13 | Matt Pocock skills | 用于真实工程的小型可组合技能 | ~224k |
| 14 | SkillClaw | 在实际使用后自我改进的技能 | ~2.5k |
| 15 | Browser Harness | 智能体驱动你的真实浏览器并学习辅助函数 | ~17k |
| 16 | codebase-memory-mcp | 将代码仓库索引为知识图谱供智能体使用 | ~40k |
| 17 | Loopy / Loop Library | 带检查与停止规则的可重复智能体循环 | ~3k |
## “技能”到底是什么
把技能想象成一张食谱卡片，你的智能体在需要时可以随时取用。
没有技能时：
- 每次会话你都要重新解释同样的工作流程
- 质量取决于模型当天的心情
- 多步骤工作容易跑偏
有技能时：
- 步骤只需写下来一次
- 任务匹配时智能体自动加载
- 同一份食谱可以跨项目共享
Hermes 本身已内置许多技能。这 17 个是人们想要一套严肃配置时额外加装的外部技能包。
## 1. Agent-Reach
仓库： [Panniantong/Agent-Reach](https://github.com/Panniantong/Agent-Reach) 星标： ~73k
它是什么
一个 CLI 工具，让你的智能体能“看到”公开互联网——X/Twitter、Reddit、YouTube、GitHub 等等——无需你为每个平台接入付费 API。
解决的问题
你的智能体会写代码。但让它“查一下 Twitter 上的反响”或“总结这个 YouTube 教程”，它经常失败：爬虫被封、付费 API、登录墙、乱糟糟的 HTML。
对新手有什么帮助
安装一次，然后就可以这样说话：
- 在 X 上搜索人们对这个工具的评价
- 拉取这个 YouTube 视频的字幕并总结
- 去 Reddit 查查这个 bug
什么时候用
内容调研、竞品分析、舆情监听、技术支持调研。
## 2. make-interfaces-feel-better
仓库： [jakubkrehel/make-interfaces-feel-better](https://github.com/jakubkrehel/make-interfaces-feel-better) 星标： ~3k
它是什么
一个专注于 UI 打磨的智能体技能：动画、排版、图标、悬停状态、点击区域、阴影、视觉对齐——那些让界面显得“用心”的细节。
解决的问题
AI 构建的 UI 通常“能用”但感觉不对。间距错误。悬停没反应。按钮太小。文字一看就是自动生成的。
对新手有什么帮助
智能体建好页面后，加载这个技能，让它从“手感”而非“功能”的角度过一遍 UI。
什么时候用
落地页、仪表盘、内部工具、任何智能体刚交付的前端。
安装示例
```
npx skills add jakubkrehel/make-interfaces-feel-better
```
## 3. oh-my-hermes (OMH)
仓库： [witt3rd/oh-my-hermes](https://github.com/witt3rd/oh-my-hermes) 星标： ~270
它是什么
专为 Hermes 打造的多智能体编排技能，灵感来自 oh-my-claudecode。包含调研、访谈、共识规划和已验证执行的技能。
内部核心技能
| 技能 | 用途 |
| --- | --- |
| omh-deep-research | 拆分调研 → 并行搜索 → 综合 → 核查引用 |
| omh-deep-interview | 苏格拉底式需求访谈 |
| omh-ralplan | 规划者/架构师/批评者辩论，直到计划站得住脚 |
| omh-ralph | 实现 → 验证 → 循环直到完成 |
| omh-autopilot | 以上流程的完整管线 |
解决的问题
一个智能体自由发挥大任务时，常常跳过需求、规划草率，然后没有证明就宣称“完成”。
什么时候用
陌生领域、跨多天的功能开发、任何你希望在写代码前先保证计划质量的工作。
建议流程
```
research → interview → ralplan → ralph
```
## 4. Anthropic Cybersecurity Skills
仓库： [mukul975/Anthropic-Cybersecurity-Skills](https://github.com/mukul975/Anthropic-Cybersecurity-Skills) 星标： ~30k
它是什么
一个面向智能体的大型开源结构化网络安全技能库（约 817 个技能，覆盖多个安全领域）。映射到 MITRE ATT&CK、NIST CSF、ATLAS、D3FEND、AI RMF 等框架以及欺诈防范框架。
重要的诚实说明
这是一个社区项目，不是 Anthropic 官方产品。它包含攻击性/双用途技术，仅限用于授权测试、防御、研究和教育。
解决的问题
你的编程智能体默认并不是资深安全分析师。这些技能为它提供了调查、检测、取证式工作流以及与框架对齐的指导手册。
什么时候用
安全审查、事件响应草案、威胁建模、授权红队演练准备、跟着智能体导师学习安全工作流。
新手规则
只对你拥有的系统或获得书面许可测试的系统使用。
## 5. OpenMontage
仓库： [calesthio/OpenMontage](https://github.com/calesthio/OpenMontage) 星标： ~49k
它是什么
一个开源的智能体视频制作系统。管线、工具和制作知识文件，能把编程智能体变成接近一个小型视频团队的存在。
解决的问题
视频工作是一堆工具的堆叠：脚本、B-roll 素材、剪辑、字幕、导出。大多数智能体只能“聊聊视频”。OpenMontage 瞄准的是实际制作工作流。
什么时候用
短内容、解说视频、产品演示、智能体参与其中的 YouTube 式制作。
适合谁
每天都在使用智能体、想要媒体产出而不仅仅是文字的创作者和营销人员。
## 6. Mission Control / Minions
仓库： [agent37-platform/minions](https://github.com/agent37-platform/minions) 星标： ~625
它是什么
一个看板式的 Hermes 工作控制面板。创建任务、观看运行、审查结果，让长任务不会消失在终端历史里。
解决的问题
Hermes 本身很强，但在裸终端里管理五个并行会话就不行了。你会忘记什么完成了、什么卡住了、什么在等你批准。
重要功能
- 看板：进行中 / 待审查 / 已完成
- 工具调用的实时流
- “完成”前需要人工签核
- 每个任务单独的模型设置
- 定时 Hermes 任务
- 本地优先选项（SQLite，数据留在你的机器上）
快速开始
```
npx minionsai
```
然后打开 http://localhost:6969。
什么时候用
任何把 Hermes 当成真正的干将、而不是一次性聊天玩具来用的人。
## 7. Resemble AI Detect
仓库： [resemble-ai/detect-skill](https://github.com/resemble-ai/detect-skill)
它是什么
一个媒体安全智能体技能——使用 Resemble 的检测技术栈来检测 AI 生成或篡改的音频、图像和视频。
解决的问题
深度伪造和合成媒体无处不在。如果你的智能体要处理 UGC、品牌安全或调研工作，在信任一段素材之前你需要一个检查步骤。
什么时候用
内容审核管线、品牌安全审查、在引用或转发前检查“这段素材是真的吗？”
诚实说明
把检测分数当作信号，而不是法庭证据。高风险判断永远保留一个人工裁决步骤。
## 8. addyosmani/agent-skills
仓库： [addyosmani/agent-skills](https://github.com/addyosmani/agent-skills) 星标： ~89k
它是什么
来自 Addy Osmani 世界的生产级工程技能：定义 → 计划 → 构建 → 验证 → 审查 → 发布的完整生命周期。
命令心智模型
| 命令 | 含义 |
| --- | --- |
| /spec | 定义要构建什么 |
| /plan | 拆分为小任务 |
| /build | 分片实现 |
| /test | 证明它能用 |
| /review | 改善代码健康度 |
| /webperf | 先测量再优化 |
| /ship | 小心地把它上线 |
解决的问题
毫无关卡凭感觉写代码。这些技能强制执行资深工程师已有的习惯：规格说明、小任务、测试、审查。
什么时候用
真正的产品，而不是一次性的脚本。尤其适合前端/Web 工作和跨多天的功能开发。
安装示例
```
npx skills add addyosmani/agent-skills
```
## 9. Composio skills
仓库： [ComposioHQ/skills](https://github.com/ComposioHQ/skills)
它是什么
可分发的技能，教会智能体如何使用 Composio 构建——工具路由、认证、会话、触发器、生产级模式。
解决的问题
把智能体连接到 Gmail、Slack、GitHub、CRM 等服务很麻烦。Composio 是集成层；这些技能教给智能体安全的模式（用户 ID、会话、Webhook、框架接线）。
什么时候用
任何需要操作真实应用、而不仅仅是本地文件的 Hermes/智能体配置。
安装示例
```
npx skills add composiohq/skills
```
## 10. youtube-full（YouTube 技能）
仓库： [ZeroPointRepo/youtube-skills](https://github.com/ZeroPointRepo/youtube-skills) 星标： ~540
它是什么
让你的智能体能够：
- 拉取 YouTube 字幕
- 搜索视频
- 浏览频道
- 提取播放列表
而不用与云 IP 上的 yt-dlp 封锁搏斗。
解决的问题
“总结这个 YouTube 视频”是智能体最常见的失败场景之一。这个包就是为这个任务而生的。
Hermes 安装示例
```
hermes skills install skills-sh/ZeroPointRepo/youtube-skills/skills/youtube-full
```
什么时候用
从教程学习、内容调研、竞品视频拆解、课程笔记。
免费额度说明
底层字幕 API 宣称注册赠送免费额度。请在其官网确认当前限额。
## 11. Humanizer
仓库： [blader/humanizer](https://github.com/blader/humanizer)
它是什么
一个重写 AI 味文字的技能，让文本读起来像人写的。基于 Wikipedia 的“AI 写作迹象”模式。
解决的问题
你的智能体草稿很有用，但读起来仍然像新闻通稿：“关键性的”、“格局”、“这不是 X，而是 Y”、完美的三段式列表、空洞的吹嘘。
实际工作方式
1. 先不保护 AI 结构进行重写
2. 检查剩余的 AI 痕迹
3. 再次重写
4. 保留事实；不要编造名字/数字
什么时候用
智能体写的每一篇公开帖子、文档页面、客户邮件或发布说明。
专业技巧
粘贴两三段你自己写的文字，让它匹配你的语调。
## 12. Defuddle
仓库： [kepano/defuddle](https://github.com/kepano/defuddle) 星标： ~9k
它是什么
一个库/CLI，输入乱糟糟的网页，返回以干净的 HTML 或 Markdown 呈现的正文内容。由 Obsidian Web Clipper 圈子（kepano）打造。
解决的问题
智能体抓取网页时经常被导航栏、广告、Cookie 横幅和嵌套 HTML 噎住。Defuddle 把垃圾统统剥掉。
什么时候用
调研、把文章剪藏进笔记、给 Hermes 提供干净的上下文、构建你自己的阅读器管线。
简单的心智模型
输入 URL 或 HTML → 输出可读的文章。
## 13. Matt Pocock skills
仓库： [mattpocock/skills](https://github.com/mattpocock/skills) 星标： ~224k
它是什么
Matt 的日常工程技能：小巧、可组合、可编辑。为真实应用而建，不是“凭感觉写代码的表演”。
一句话哲学
大型流程框架会夺走你的控制权。这些技能保持小巧，方便你魔改。
解决的问题
你的智能体要么没有流程，要么有一套你无法调试的巨型流程。Matt 的这套正好处于中间。
什么时候用
TypeScript/应用工程、工单分类、规划，以及与智能体一起日常交付。
安装示例
```
npx skills@latest add mattpocock/skills
```
然后每个仓库运行一次 setup 技能。
## 14. SkillClaw
仓库： [AMAP-ML/SkillClaw](https://github.com/AMAP-ML/SkillClaw) 星标： ~2.5k
它是什么
一个让技能从真实使用中演进的系统。你正常聊天。它在后台消化会话、改进技能、去重垃圾，并在多个智能体/设备之间积累经验。
解决的问题
你的 Hermes 技能文件夹会变成杂物抽屉：重复项、半成品笔记、过时的步骤。没有人“消化”这些学习成果。
什么时候用
如果你每天都在 Hermes 里工作，希望技能库随时间越来越锋利，而不用每周手动维护。
诚实定位
这是面向高级用户的基础设施。等你已有了一套基础技能后再用才划算。
## 15. Browser Harness（荣誉提名）
仓库： [browser-use/browser-harness](https://github.com/browser-use/browser-harness) 星标： ~17k
它是什么
一个自我修复的浏览器套件。LLM 通过 CDP 连接到你的真实浏览器，执行任务，并在某个步骤尚不支持时自动编写缺失的辅助函数。
解决的问题
静态浏览器工具在网站一变就挂掉。这种方式可以边工作边改进套件。
示例任务
“打开我的 X 主页，抓取我最近 20 条视频帖子，把它们下载下来。”
什么时候用
已登录的浏览器工作、重复性的网页琐事、任何需要真实会话 Cookie 栈的任务。
注意
把你的真实浏览器交给智能体既强大又有风险。实验时先用一个独立的浏览器配置文件。
## 16. codebase-memory-mcp
仓库： [DeusData/codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp) 星标： ~40k
它是什么
一个高性能 MCP 服务器，将你的代码库索引为持久化知识图谱。智能体直接查询结构，而不是每次会话重读 40 个文件。
为什么重要
逐文件探索既烧 token 又容易遗漏架构。图查询可以直接回答：
- 这个在哪里被使用？
- 什么调用了什么？
- 我改了这里会破坏什么？
宣称的能力
- 158 种语言
- 极快的索引速度
- 相比朴素探索大幅节省 token
- 本地处理（代码留在你的机器上）
什么时候用
中大型仓库、多服务代码库、任何上下文折腾让你抓狂的 Hermes 编程工作。
## 17. Loop Library / Loopy（荣誉提名）
仓库： [Forward-Future/loopy](https://github.com/Forward-Future/loopy) 星标： ~3k
它是什么
两部分组成：
1. 一个公开的智能体循环目录
2. Loopy，一个帮助查找、审计、适配、编写和运行这些循环的技能
“循环”是什么
一次性提示词说：做一次这件事。循环说：做一步，测量，只有有效才保留，重复直到满足停止条件。
示例：
> 找到最慢的页面，改进一点，再测量一次。只有有效才保留改动。达到目标或收益消失时停止。
什么时候用
性能优化、不稳定的测试、文档清理、周期性运维——任何第一次尝试永远不会是最终答案的工作。
核心思想
循环之所以强大，是因为它们是有边界的。清晰的检查。清晰的停止。需要判断时交还给人。
## 这 17 个如何组合（简单技能栈）
入门 Hermes 栈
1. Agent-Reach —— 互联网访问
2. youtube-full —— 视频学习
3. Defuddle —— 干净的网页文本
4. Humanizer —— 干净的公开写作
5. Minions —— 在看板上总览所有任务
严肃编程栈
1. codebase-memory-mcp —— 不再重读整个仓库
2. addyosmani/agent-skills 或 Matt Pocock skills —— 工程流程
3. oh-my-hermes —— 规划/验证的多智能体流程
4. make-interfaces-feel-better —— 最后过一遍 UI
内容/媒体栈
1. OpenMontage —— 视频制作
2. youtube-full —— 素材来源
3. Humanizer —— 字幕和帖子
4. Resemble Detect —— 媒体真伪检查
高级用户成长栈
1. SkillClaw —— 技能随真实使用而改进
2. Loopy —— 把重复工作变成循环
3. Browser Harness —— 真实浏览器琐事
4. Composio skills —— 应用集成
## 如果你是新手，按这个顺序安装
不要第一天就装满 17 个。
第 1 天
- Agent-Reach
- Humanizer
- youtube-full
第 2 天
- Minions（如果你运行多个 Hermes 任务）
- Defuddle
- 一个工程技能包（Matt 或 Addy 的）
第 3 天及以后
- 真实仓库再上 codebase-memory-mcp
- 任务变大时再上 oh-my-hermes
- OpenMontage / Browser Harness 只在需要那些工作时再装
技能不是越多越好。一小套你真正会触发的技能，胜过一个杂物抽屉。
## 安全与常识
其中一些会接触真实浏览器、真实代码库或安全技术。
简单规则：
- 只在你被允许测试的系统上运行攻击性安全技能
- 为自主浏览器智能体使用独立的浏览器配置文件
- 把安装脚本管道进 shell 之前先读一遍
- 不要把密钥放进提示词和技能日志
- 数据敏感时优先选择本地优先的工具
配齐这套技能的全新 Hermes 智能体是一个聪明的系统：
- 无需伺候 API 的互联网调研
- 不会折腾 token 的仓库记忆
- 不会凭感觉把你的应用写进坑里的工程流程
- 能看视频、能润色草稿的内容工具
- 一个让十项任务不会消失的看板
- 让整套配置随时间越来越锋利的循环与自我进化技能
如果我在一台新机器上只有一个小时，我会按这个顺序安装：
1. Agent-Reach
2. Humanizer
3. youtube-full
4. Matt Pocock skills 或 addyosmani/agent-skills
5. codebase-memory-mcp
6. Minions
然后我会先交付一周的真实工作，再去碰其余的。
不要像收集贴纸一样收集技能。只安装那些能消除你本周痛点的技能。
上不了网 → Agent-Reach 假味草稿 → Humanizer 会话混乱 → Minions 仓库折腾 → codebase-memory-mcp 工程流程薄弱 → Matt 或 Addy 视频工作 → OpenMontage 真实浏览器琐事 → Browser Harness
这就是我会信任的全新配置
如果这篇文章对你有帮助，请给个关注。我每周都会发布免费 API 和智能体工作流。
## 关注 [@fareanfts](https://x.com/@fareanfts) 获取更多 🙌
## 相关链接
- [Farea](https://x.com/FareaNFts)
- [@FareaNFts](https://x.com/FareaNFts)
- [25K](https://x.com/FareaNFts/status/2091538403611533820/analytics)
- [Panniantong/Agent-Reach](https://github.com/Panniantong/Agent-Reach)
- [jakubkrehel/make-interfaces-feel-better](https://github.com/jakubkrehel/make-interfaces-feel-better)
- [witt3rd/oh-my-hermes](https://github.com/witt3rd/oh-my-hermes)
- [mukul975/Anthropic-Cybersecurity-Skills](https://github.com/mukul975/Anthropic-Cybersecurity-Skills)
- [calesthio/OpenMontage](https://github.com/calesthio/OpenMontage)
- [agent37-platform/minions](https://github.com/agent37-platform/minions)
- [resemble-ai/detect-skill](https://github.com/resemble-ai/detect-skill)
- [addyosmani/agent-skills](https://github.com/addyosmani/agent-skills)
- [ComposioHQ/skills](https://github.com/ComposioHQ/skills)
- [ZeroPointRepo/youtube-skills](https://github.com/ZeroPointRepo/youtube-skills)
- [blader/humanizer](https://github.com/blader/humanizer)
- [kepano/defuddle](https://github.com/kepano/defuddle)
- [mattpocock/skills](https://github.com/mattpocock/skills)
- [AMAP-ML/SkillClaw](https://github.com/AMAP-ML/SkillClaw)
- [browser-use/browser-harness](https://github.com/browser-use/browser-harness)
- [DeusData/codebase-memory-mcp](https://github.com/DeusData/codebase-memory-mcp)
- [Forward-Future/loopy](https://github.com/Forward-Future/loopy)
- [@fareanfts](https://x.com/@fareanfts)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [10:49 PM · Aug 23, 2026](https://x.com/FareaNFts/status/2091538403611533820)
- [25K Views](https://x.com/FareaNFts/status/2091538403611533820/analytics)
---
*导出时间: 2026/8/24 14:30:22* 
