为什么值得关注

它把每个客户的记忆、销售框架的 RAG 检索和严格的「反 AI 味」写作规则结合在一起，而且你只需在 Telegram 里回答七个问题就能教会它你卖什么，不用改配置文件。

适合谁使用

在 Telegram DM 里做销售的自由职业者或创始人
需要实时回复建议和客户跟进的顾问
愿意自托管、重视数据隐私的技术型销售团队

典型使用场景

把潜在客户的最新消息粘贴到主题中，获得策略分析与三条可直接发送的回复
用独立主题管理多个潜在客户，记录阶段、温度、痛点和预算信号
通过 /setup 或 /import 让机器人了解你的产品，并用 /profile add 持续补充事实
把自有销售话术或书籍加入知识库并重新导入，提升建议质量

项目优势

每个潜在客户对应一个论坛主题，完整历史进入每次提示，机器人能追踪漏斗阶段
内置 14 个销售框架种子文件（SPIN、Voss、Cialdini、Hormozi 等），也支持加入自己的资料
硬性反 AI 写作规则：避免破折号、排比、「Hope this helps」等机器人腔
自托管、两个 SQLite 文件、无需公网端口；默认配置每次回复成本约 5-6 美分，并有成本估算脚本

使用前须知

需要自己的 Anthropic API 密钥，并承担按次使用成本；RAG 还需要 OpenAI 密钥
依赖 Telegram 论坛主题，且机器人必须拥有群组管理员权限，否则收不到普通消息
如果跳过 /setup 或 /import，回复会包含 [your price range] 之类的占位符；机器人质量取决于你教给它的画像

README 快速开始

DM Sales Coach

Русская версия →

A self-hosted Telegram bot that coaches you through sales conversations in DMs, in real time.

You paste what a prospect just wrote you. It replies with a read of what's actually going on, three copy-paste-ready reply options using different strategies, and the next move. Each prospect lives in its own Telegram forum topic, so the bot remembers the whole thread and tracks where they are in your funnel.

It learns what you sell by interviewing you in Telegram — seven questions, answered the way you talk. No company is baked into this repo, and you never have to edit a config file to teach it yours.

You  ──▶  [paste what the prospect said, into their topic]

Bot  ──▶  Arc: 2/4, short arc, warming fast
          Read: he's price-anchoring before scope. Answering now caps your ceiling.

          A — Reframe:        depends what we'd build. quick q: validating this, or scaling
                              something that already works?
          B — Anchor high:    most builds land $3-15k. giving you a number before I
                              understand the workflow would be guessing.
          C ★ — Light qualifier:  ballpark $3-15k. what's the actual task eating the
                              hours right now?

          Next: hold the line on no-price-before-scope. If he pushes again, he's qualified.

          📌 temp: 7  •  pain: manual invoice routing

Why it works

Three things, and the third is the one people skip:

Per-prospect memory. Every prospect is a forum topic. Full history goes into every prompt, so advice on message 5 knows what happened on message 1.
Retrieval over a sales knowledge base. Ships with 14 seed files covering SPIN, Voss, Cialdini, Hormozi, objection handling, pricing conversations, and DM outreach mechanics. Add your own books and saved DM transcripts on top.
Hard anti-AI writing rules. No em dashes, no "not just X — it's Y", no triads, no "Hope this helps." The drafts are meant to look like a human typed them on a phone, because a prospect can smell an LLM in a DM.

Stack

aiogram 3 — Telegram bot framework
Claude — coaching generation, with prompt caching on the static prefix
OpenAI text-embedding-3-small — embeddings only, optional
SQLite + [sqlite-vec](https

项目描述

Self-hosted Telegram bot that coaches you through sales conversations in DMs. Interviews you about your business, then drafts replies that sound human. Claude + RAG, one process, your data stays yours.

dm-sales-coach