automation
How to write a system prompt for a Telegram AI sales bot
A vague system prompt turns your Telegram AI sales bot into a liability. Here's the exact structure that keeps it on-script and closing.
Your Telegram AI sales bot just told a prospect your product does something it doesn't. Now you're on a call explaining why the "AI" promised a refund policy that doesn't exist. This isn't a bug — it's a prompt problem.
Most bots that go off the rails weren't broken by bad AI models. They were broken by a system prompt that gave the model too much room to improvise and not enough structure to stay on task.
What should a system prompt for a Telegram sales bot actually include?
A working system prompt needs 6 core components: a defined role, a strict scope of what it can and can't discuss, a conversation goal, tone/voice rules, escalation triggers, and reference to a knowledge base for product-specific facts. Skip any one of these and the bot will eventually improvise its way into a mistake — quoting a price that changed, promising a feature that doesn't exist, or arguing with a frustrated prospect instead of handing off.
The goal isn't to write a clever prompt. It's to write a boring, specific one that leaves as little as possible to the model's imagination.
Start with the role and the single goal
Every good sales prompt opens by telling the bot exactly who it is and what winning looks like. Not "be helpful" — a specific, measurable outcome.
Define the persona: Name, role, company, and tone in one sentence ("You are Alex, a sales assistant for [Company], friendly but direct, never pushy").
State the one goal: Book a call, get a form filled out, or push a link — pick one primary action, not three.
Set the fallback: Tell it explicitly what to do when it doesn't know the answer — hand off, don't guess.
List forbidden topics: Pricing exceptions, legal claims, competitor comparisons — anything that needs a human sign-off.
Cap the conversation length: If it's gone 8-10 messages without progress toward the goal, it should escalate rather than loop.
This is how the Telegram AI Sales Agent is designed to work in practice — it sits in your Telegram account, replies to incoming messages according to your prompt, and either drives the conversation toward an action like booking a meeting or filling a form, or pauses for your intervention when the question falls outside what it's confident answering.
Give it a knowledge base instead of stuffing facts into the prompt
Don't cram your entire product catalog into the system prompt. It bloats every single message the model processes and makes the bot slower to respond and more likely to blend facts incorrectly.
Instead, separate the two: the system prompt handles behavior (how it talks, what it's trying to do, when to escalate), and a knowledge base handles facts (pricing tiers, FAQs, edge cases). CRMChat's setup for the AI sales agent explicitly splits these — you need a prompt and, separately, a knowledge base if your product is complex enough to need FAQs on hand. If you haven't built one yet, this guide to building a knowledge base for a Telegram support bot walks through the structure.
Write escalation rules before you write anything else
The single biggest failure mode for sales bots isn't a bad answer — it's a bad answer given with total confidence. Your prompt needs explicit rules for when the bot should stop and wait for a human.
Ambiguous intent: If the prospect's question doesn't clearly match a known FAQ, don't guess — flag for handoff.
Price or contract negotiation: Any custom discount, custom terms, or contract language goes to a human, always.
Anger or frustration signals: Words like "refund," "cancel," "scam," or repeated exclamation points should trigger immediate escalation.
Repeated questions: If the same question comes up twice in one thread, the bot isn't landing — hand it off.
Anything legal, medical, or financial: Bots shouldn't be making claims in regulated categories, full stop.
Once you've defined those triggers, make sure the handoff itself is smooth — a bot that just goes silent looks worse than one that never replied. This piece on handing off a Telegram AI conversation to a human agent covers how to make that transition feel intentional instead of broken.
Test it like a skeptical prospect, not like the person who built it
The prompt that looks perfect in your head will break the first time a real person asks something sideways. Before you launch, run it through adversarial testing yourself:
Ask it the same question three different ways and check for consistent answers
Try to get it to quote a price or promise it shouldn't make
Act frustrated and see if it escalates correctly
Ask something completely off-topic and confirm it redirects instead of engaging
Check response length — sales prospects on Telegram don't want a paragraph, they want two sentences
If you're building this for a client rather than your own team, it's worth running the whole thing through a proper staging environment before launch so you catch these failures before the prospect does.
Use a prompt-writing template instead of starting from a blank page
Writing a system prompt from scratch every time is how inconsistency creeps in. CRMChat points to a dedicated GPT-based instruction assistant built specifically for writing these prompts — it walks you through role, goal, tone, and escalation logic in a structured Q&A instead of you guessing at format. It's a faster starting point than a blank text box, especially if you're setting up bots for multiple clients or products.
Once the prompt is drafted, plug it into the bot and pair it with your knowledge base. From there, the bot handles the conversation according to the prompt — answering questions, nudging toward the goal, and escalating when it hits the boundaries you set.
Keep the account healthy while the bot runs
A great prompt doesn't matter if the Telegram account running it gets flagged for spam-like behavior. Bots that message too fast or too generically can trigger the same account restrictions as manual spam outreach. If you're deploying this on an outreach account rather than an inbound support line, pair your prompt work with proper account warmup so the account has the history to support bot-driven conversations at scale.
If you want to see the full setup — including how the agent handles handoffs and where the knowledge base plugs in — the CRMChat Help Center has the walkthrough with screenshots.
