全部使用案例
销售
网站访客 → 外联:先写清规格,再让 Grok Bot 搭建
手搓 Clay 表和访客→邮件管道吃掉一周;含糊需求只会得到含糊系统。
当时卡在哪
@BenyaminHolley 让 Grok Bot 搭常青访客→外联栈(Clay Web Intent + Squid、Slack 提醒、富化瀑布、按页面写文案),以前在 AirOps 手搓要一周。最大教训:先写终态、硬约束和完成长什么样。回复里有完整可粘贴 prompt——含每个 PR 合并前给你看、开启发信前等你放行。tip + 作者原文 prompt。与 xAI 无关。
要装什么、建什么
写规格 → 填括号粘贴 prompt → 审 PR → 试几行 → 批准上线。
机器人
不新建机器人。按「第一次怎么跑」做。
插件
GitHub
必装
Slack
可跳过
例行任务
第一次先别开。自己看过一轮再说。
提示词
常青:网站访客 → 外联工作流
引自 @BenyaminHolley — 按其公开原文。不要另编系统提示词。
最后一行是我们加的审批围栏:Never send, post, spend, or buy without my explicit yes.
(作者英文原文) Build me an evergreen website visitor to outbound workflow, end to end, with no human in the loop. My site is [domain], built on [framework/host]. Use GitHub and Cursor cloud agents for code changes, and show me every PR before merge. Website 1. Add the Clay Web Intent tracker (company-level) and the [Squid/RB2B] tracker (person-level) directly in the site code, not Tag Manager. Confirm both load on the live site. Data flow 2. Clay Web Intent writes companies into one Clay table. Only keep sessions of 20+ seconds or 2+ pages. 3. Send person-level visitors by webhook into a separate Clay table, with an auth header. 4. Have Clay POST every visitor to a Slack bot or webhook with a secret. Dedupe per person or company per day. Post pages viewed, time on site, referrer, UTMs and contacts found. Skip my own team's domain and emails. Filtering and enrichment 5. Filter out agencies, competitors, GTM vendors, gov, schools and current customers before spending credits. 6. For company visitors, use AI to pick up to 3 buyers by role. Dedupe on a normalized LinkedIn URL. 7. Email waterfall: [LeadMagic] first, then pattern guesses, and validate every one. Every lead needs a first name, last name, LinkedIn URL and a valid email. Copy 8. Tag my content library by best-fit role and industry. AI picks one piece per person. 9. Company visitors get one AI line on what the company sells, a soft offer, and the content piece in email 2. 10. Person visitors get a campaign per key page: [pricing, case studies, blog, contact with no submit, repeat visitor]. The page sets the angle. 11. Never say or imply they visited the site. Lint every row and block anything that does, or that links outside my library. Launch 12. Suppress against [Salesforce/HubSpot], then push to [EmailBison/Instantly]. 13. Test on a few real rows, show me the output, and wait for my go before turning it on. Never send, post, spend, or buy without my explicit yes.
第一次怎么跑
- 1阅读 https://x.com/BenyaminHolley/status/2116298502225420345 与 prompt 回复 https://x.com/BenyaminHolley/status/2116298504436113455(致谢 @BenyaminHolley)。
- 2打开 Bot 前先写终态、硬约束和完成标准。
- 3粘贴作者 prompt;填好 [域名]、追踪、富化、CRM/ESP 括号。
- 4Bot 开的每个 GitHub PR 合并前你审;追踪不要走 Tag Manager。
- 5在 Slack/Clay 试几行真实数据,再明确批准上线——绝不自动发信。
审批围栏
每个 PR 等你合并;外联发送保持关闭,直到你批准真实样例测试。
你该改什么
偷:规格优先 + 「别说你来过站」检查 + 外联前等人放行。
出处
benyamin @BenyaminHolley · 原帖
与 xAI 无关。提示词不是我们写的。第三方模板条款照旧适用。
Pro 设置包
想要完整设置,而不只是思路?
Pro 设置包为热门用例提供岗位描述、提示词、排程与踩坑。