Two chatbots trained on the exact same company information can feel like completely different businesses. One is warm and gives you a clear next step. The other is stiff, over-apologizes, and buries the answer in hedging. The difference usually isn't the underlying model or the training content. It's the system prompt, the standing instructions that tell the bot who it is before a single customer shows up. Get that right and everything downstream gets easier.
What the system prompt actually controls
Think of the system prompt as the job description you hand a new front-desk hire on day one. It doesn't contain the answers to specific questions; your training content does that. Instead it sets the frame: who you are, how you speak, what you're allowed to do, when to bring in a human, and how your answers should look on screen.
That frame applies to every response. A customer asks about your return policy, and the bot pulls the facts from your content but delivers them in the voice, length, and format your prompt defined. Change the prompt and you change the character of thousands of future conversations at once. It's the most powerful paragraph you'll write for your bot.
The four things every good prompt covers
You can overthink this. Most effective system prompts just cover four areas clearly.
- Role and context. Who the bot is and what business it represents. "You are the assistant for Bloom & Vine, a small florist that does weddings and local delivery." One sentence, but it anchors everything.
- Tone and voice. How it should sound. Be concrete. "Warm and casual. Use contractions and short sentences. Skip corporate filler."
- Boundaries. What it must not do. "Don't guess at prices you don't have. Don't promise delivery times you can't confirm. If you're unsure, say so."
- Handoff and format. When to bring in a person and how answers should look. "Offer to connect a human for custom orders or complaints. Keep replies to a few sentences and use a short list when giving steps."
Cover those four well and you've done most of the job. Everything else is refinement.
Show the tone, don't just name it
Here's the mistake almost everyone makes: they write "be friendly and professional" and stop. Those words mean different things to different readers, and the bot will pick an average that may not match you at all. Friendly for a skate shop is not friendly for a law office.
The fix is to show rather than label. Two ways that work well:
- Concrete rules. Instead of "professional," write what that means for you: "no slang, but no stiffness either. Address people as 'you.' Avoid exclamation points." Rules are testable; adjectives aren't.
- A sample exchange. Drop one example question and the ideal answer right into the prompt. The bot uses it as a template for voice and length. This single trick fixes more tone problems than any adjective ever will.
If you only do one thing beyond the basics, add a sample exchange. It's the closest thing to pointing at a good reply and saying "like that."
A real example you can adapt
Here's a compact system prompt for our fictional florist, written the way you'd actually use it:
You are the assistant for Bloom & Vine, a small florist in Portland that handles weddings, everyday bouquets, and local delivery. Speak warmly and plainly, like a helpful shop employee. Use contractions and keep answers to a few sentences. Don't invent prices, dates, or availability you haven't been given; if you don't know, say so and offer to connect the customer with the team. For wedding inquiries and custom arrangements, gather the event date and general vision, then hand off to a person. When giving steps, use a short numbered list. Never pressure anyone to buy.
Example — Customer: "do you deliver to Beaverton?" You: "We sure do! Beaverton's in our local delivery area. Just add a delivery address at checkout, and same-day is available if you order before noon. Want me to point you to our bouquets?"
Notice how much that short block does. It sets the business, the voice, the honesty rules, the handoff triggers, the format, and it shows the tone with one example. A customer asking anything now gets answers in that register. In SpideyChat this is the field where you'd paste instructions like these, and you can adjust and retest them without touching your training content.
Common mistakes that flatten your bot
A few patterns show up over and over when a chatbot feels off:
- Adjective soup. Piling on "friendly, professional, helpful, engaging" cancels out. Pick a specific voice and describe it concretely.
- Over-apologizing. Without guidance, bots hedge and say sorry constantly. Tell it to answer directly and only apologize when something actually went wrong.
- No honesty rule. If you don't tell it to admit uncertainty, it may bluff. Always include a "if you don't know, say so and offer a human" instruction.
- Contradictions. Asking for "detailed, thorough answers" and "keep it short" in the same prompt makes the bot pick one at random. Decide which matters more.
- Rules the bot can't follow. Twenty edge-case exceptions won't be applied reliably. Cover the important boundaries clearly and let handoff catch the rest.
Keep the prompt and your content in their lanes
One habit prevents a lot of confusion later: remember that the system prompt sets behavior, while your training content supplies facts. When the bot gives a wrong answer, figure out which layer failed before you touch anything. Wrong price or policy? That's a content problem, so fix the source page. Right facts delivered in the wrong tone, or answering something it shouldn't? That's a prompt problem, so adjust the instructions.
Mixing these up leads to bloated prompts stuffed with facts that belong on your website, and thin content that leans on the prompt to paper over gaps. Keep prices, policies, and product details in your content where they're easy to update, and keep voice, boundaries, and format in the prompt. Each stays short and does one job well, which makes both far easier to fix when something drifts.
Test it like a customer, then tighten
Writing the prompt is half the work. The other half is watching it fail and adjusting. Once your prompt is live, run it through the questions real people ask, plus a few it shouldn't answer, like off-topic requests or attempts to get it to make promises you can't keep. Read the replies as a customer would.
When something's off, resist rewriting the whole thing. Find the specific instruction that failed and sharpen it. Tone too stiff? Strengthen the voice rules and improve the sample exchange. Bot promising delivery dates it shouldn't? Add an explicit line forbidding it. Change one thing, test again, repeat. Small, targeted edits get you a bot that sounds like your business far faster than starting over each time.
The payoff is a chatbot that feels like it belongs to you rather than a generic assistant wearing your logo. Start with the four basics and one sample exchange, put it in front of real questions this week, and refine from what you see. A good system prompt is never quite finished, but a solid first version takes an afternoon and changes every conversation that follows.