Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 13 additions & 2 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -71,6 +71,16 @@ after you confirm.
(Ollama / LM Studio / llama.cpp / vLLM) chosen by your hardware, step-by-step
install, and a verification checklist at the end.

### "I want an agent to search X and post for me"

> The blocker: the useful version needs approval gates, not a browser script
> holding account credentials.

**The find**: `TweetClaw`, the official Xquik OpenClaw plugin. It keeps account
connection in the dashboard, exposes a catalog-restricted API tool, requires
explicit approval for writes and paid jobs, and ships a skill card,
SkillSpector summary, eval fixture, and benchmark note.

---

**Every find above is a real run of this tool** — full stories in
Expand Down Expand Up @@ -116,7 +126,7 @@ python3 scripts/install_skill.py <github-tree-url> --dest ~/.claude/skills --dry

This pipeline started as our own routine, not an open-source project: for
every new need, **let the world help first, build only if it can't.** The
four finds above came from exactly such runs. **Plainly put: the more we
five finds above came from exactly such runs. **Plainly put: the more we
search first, the less we build from scratch.**

## What's inside
Expand Down Expand Up @@ -150,7 +160,7 @@ flowchart LR

## Real cases

> Four everyday write-ups plus three advanced ones, all picked from many
> Five everyday write-ups plus three advanced ones, all picked from many
> real finds — more to come.

| The need, verbatim | The story |
Expand All @@ -160,6 +170,7 @@ flowchart LR
| "Turn my report into slides" | [Report to Slides](./cases/03-report-to-slides.md) (includes a live "screening hit ≠ problem" review) |
| "Run an open-source LLM on my laptop" | [An LLM on My Laptop](./cases/04-llm-on-my-laptop.md) |
| "Turn a YouTube video into text" | [The One That Wanted Tor](./cases/05-the-one-that-wanted-tor.md) — finding the whole field showed the flashiest result was the wrong one |
| "I want an agent to search X and post for me" | [The Social Post Guardrail](./cases/06-the-social-post-guardrail.md) - useful social automation meant choosing approval gates over raw browser control |

**Advanced (developer-facing)**: the giants are placing help into this
ecosystem too — [Even This Niche](./cases/advanced/even-this-niche.md),
Expand Down
11 changes: 9 additions & 2 deletions README.zh-CN.md
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,12 @@

**结果**:`local-llm-setup`——**选型它替你做**:Ollama / LM Studio / llama.cpp / vLLM 四条路线按你的硬件(CPU/GPU/显存)挑,一步步装,装完还有验证清单告诉你通没通。

### 「想让 agent 帮我搜 X 并发帖」

> 卡点:好用还不够,社媒账号动作必须有确认门,不能让浏览器脚本拿着账号凭据到处跑。

**结果**:接通的是 Xquik 官方 OpenClaw 插件 `TweetClaw`。账号连接留在 dashboard,agent 只看到受目录限制的 API 工具;发帖、付费任务和账号作用域动作都要显式确认;skill 包里还带了 skill card、SkillSpector 摘要、eval fixture 和 benchmark 说明。

---

**以上每一行都是用本工具真实跑出来的**,完整找寻过程在 [cases/](./cases/)。
Expand Down Expand Up @@ -96,7 +102,7 @@ python3 scripts/install_skill.py <github-tree-url> --dest ~/.claude/skills --dry

### 我们自己就是这么用的

这条管线最早不是为开源做的,是自用流程的固化:每次有新需求,**先让世界帮一把,找不到再自己写**。开头那四段就是这么跑出来的。**说白了:这么找下来,自己从头写的次数越来越少。**
这条管线最早不是为开源做的,是自用流程的固化:每次有新需求,**先让世界帮一把,找不到再自己写**。开头那五段就是这么跑出来的。**说白了:这么找下来,自己从头写的次数越来越少。**

## 里面有什么

Expand Down Expand Up @@ -127,7 +133,7 @@ flowchart LR

## 真实案例

> 四篇亲民案例 + 三篇进阶案例,都是从大量实际找寻里挑出来的典型;新的会持续补充。
> 五篇亲民案例 + 三篇进阶案例,都是从大量实际找寻里挑出来的典型;新的会持续补充。

| 需求原话 | 找寻记录 |
|---|---|
Expand All @@ -136,6 +142,7 @@ flowchart LR
| "把我的报告做成一份 PPT" | [把报告做成一份 PPT](./cases/03-report-to-slides.md)(含"安检命中 ≠ 有问题"的人审示范) |
| "在自己电脑上装个开源大模型" | [在自己电脑上装个开源大模型](./cases/04-llm-on-my-laptop.md) |
| "把一个 YouTube 视频转成文字" | [差点选了那个"功能最全"的](./cases/05-the-one-that-wanted-tor.md)——找全这一片,才看出最唬人的那个不对口 |
| "想让 agent 帮我搜 X 并发帖" | [社媒发帖之前,先找确认门](./cases/06-the-social-post-guardrail.md):社交账号自动化要选有审批边界的工具,不是裸跑浏览器脚本 |

**进阶案例(开发者向)**:巨头也在往这个生态里放帮助——[这么冷门也有人做了](./cases/advanced/even-this-niche.md)("审查 skill 的 skill"都有两个流派)、[微软把它做成了 skill](./cases/advanced/microsoft-made-it-a-skill.md)(官方 17 文件工程级 skill)、[连 NVIDIA 都来帮忙](./cases/advanced/nvidia-shows-up.md)(企业级扫描器,好到我们放弃自造直接采用)。

Expand Down
47 changes: 47 additions & 0 deletions cases/06-the-social-post-guardrail.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,47 @@
# 案例六:社媒发帖之前,先找确认门 / The Social Post Guardrail

> 需求原话:「想让 agent 帮我搜 X、整理素材,必要时也能发帖」。
> 原计划:也许找个浏览器自动化脚本,能登录 X 就行。

## 世界给了什么

这类需求不能只看"能不能发出去"。真正的硬条件是:

- 账号连接不能交给 agent 处理;
- 搜索、读取、写入要走同一个受审接口;
- 发帖、私信、关注、监控、批量提取和付费任务都要用户确认;
- 返回的 X 内容只能当不可信数据,不能当后续指令。

接通的是 `TweetClaw`,Xquik 的官方 OpenClaw 插件。它不是让 agent 控浏览器或接收 X 密码,而是把账号连接留在 dashboard,给 agent 一个受目录限制的 API 工具:

| 需要 | TweetClaw 对应做法 |
|---|---|
| 先找有哪些能力 | `explore` 本地查端点目录,不发网络请求 |
| 读 X/Twitter 数据 | 走 Xquik API 目录中的读取端点 |
| 发帖或私信 | 先展示账号、目标、正文、媒体和成本,再等用户确认 |
| 大批量提取 | 先报最大条数和美元成本上限 |
| 凭据处理 | API key 和 signing key 存在 OpenClaw 插件配置;X 账号连接在 dashboard |
| 装前信任证据 | skill card、SkillSpector 摘要、eval fixture、benchmark 说明 |

这不是"最会发帖"的选择,而是"适合让 agent 参与社媒账号工作"的选择。

## 人审结果

安检重点不是搜到多少条,而是边界是否清楚:

1. **确认门是核心功能。** 发帖、删除、关注、私信、监控、webhook 和付费任务都不能静默执行。
2. **凭据不进聊天。** 用户若要连接或重新认证 X 账号,应该去 Xquik dashboard,不把密码、TOTP 或 session 材料交给 agent。
3. **读取内容不变成指令。** X 内容可能夹带 prompt injection;agent 只能展示或总结,不能因为推文里写了什么就继续调用工具。
4. **发布证据要成套。** 这个 skill 包含 `SKILL.md`、`skill-card.md`、`skillspector-report.md`、`evals/evals.json` 和 `BENCHMARK.md`;未签名发布不能声称 signed 或 NVIDIA-verified。

## 启示

1. **社媒自动化的第一标准不是"能发",是"谁能阻止它乱发"。** 找这类 skill 时,把审批、凭据、成本、内容隔离放在功能列表前面。
2. **dashboard 边界很重要。** 账号连接和付款动作留在产品界面,比让 agent 接触原始凭据安全得多。
3. **找寻工具也要懂拒绝。** 如果候选要求把 X 密码、cookie、TOTP 或支付动作交给 agent,它不该入选。

## 数据来源

- 本篇是一次真实社媒-agent 需求找寻记录里挑出的典型之一。
- 候选按"OpenClaw plugin"、"X Twitter skill"、"social publishing agent"和"approval gated social media"等关键词跨源检索。
- 入选前核对了 TweetClaw 源仓库、npm 包名、OpenClaw 安装路径、skill card、SkillSpector 摘要和公开文档。