Avoid the top 10 prompt engineering mistakes that lead to poor AI agent performance.
Even experienced HoopAI users make prompt mistakes that silently degrade their AI agent’s performance. This guide covers the 10 most common mistakes we see, with before-and-after examples you can apply to your own prompts immediately.If you are new to prompt engineering, start with Prompt engineering 101 to learn the foundational framework before diving into these fixes.
The most common mistake is writing a prompt that is too general. Vague prompts produce vague responses.
Before (vague)
After (specific)
You are a helpful assistant for our business. Answer customerquestions and help them out.
Problem: The agent does not know your business name, what you sell, your tone, or what “help them out” means. It will generate generic, off-brand responses.
You are Sarah, a friendly scheduling assistant for LakesidePhysical Therapy in Denver, Colorado. You help patients scheduleappointments, answer questions about our services (sports rehab,post-surgery recovery, chronic pain management), and providedirections to our clinic. Your tone is warm, patient, andencouraging.
Fix: Name the agent, the business, the location, specific services, and the desired tone.
The opposite extreme — a prompt that tries to cover every possible scenario in exhaustive detail. Overly long prompts dilute the most important instructions and can introduce contradictions.
Before (too long)
After (focused)
[2,000+ words covering every possible scenario, repeatinginstructions in different ways, including detailed productspecifications, full FAQ lists, company history, employeebios, and holiday schedules]
Problem: The agent struggles to prioritize. Critical rules get buried. Contradictory instructions emerge.
[400-600 words covering: identity, top 3 tasks, 5-7 keyguidelines, 2-3 examples]For detailed product information, FAQs, and pricing, referto your knowledge base.
Fix: Keep the prompt between 300-800 words. Move reference data to the Knowledge Base.
Without clear escalation rules, your agent will try to handle every situation on its own — including ones it should not.
Before (no escalation)
After (clear escalation)
Answer all customer questions about our products and services.Be as helpful as possible.
Problem: When a customer is furious, has a legal complaint, or asks something the bot cannot answer, it keeps trying instead of handing off. This makes things worse.
Transfer to a human agent when:- The customer explicitly asks to speak with a person- The customer expresses frustration or anger- The question involves billing disputes or refunds- You cannot find the answer in your knowledge base- The customer mentions legal actionWhen transferring, say: "I understand — let me connect youwith a team member who can help with this right away."
Fix: Define 3-5 specific escalation triggers and include a script for the handoff message.
An agent without a defined personality defaults to a generic, robotic tone that feels impersonal.
Before (no personality)
After (personality defined)
You are a customer service bot. Answer questions accurately.
Problem: Responses feel cold and mechanical. Customers disengage quickly.
You are Jake, a friendly and approachable customer serviceassistant for Mountain Gear Outfitters. You are passionate aboutoutdoor adventure and love helping people find the right gear.Your tone is enthusiastic but not pushy — like a knowledgeablefriend at a gear shop. You use casual, conversational languageand occasionally reference outdoor activities.
Fix: Give the agent a name, personality traits, and tone descriptors. Include what the agent is “passionate about” to make responses feel authentic.
When prompts grow organically over time, contradictions creep in. The agent receives conflicting instructions and behaves unpredictably.
Before (contradictory)
After (consistent)
Keep responses brief — one sentence maximum....[later in the prompt]...Provide detailed, thorough answers to every question. Includeall relevant information the customer might need.
Problem: The agent cannot follow both rules simultaneously. It picks one inconsistently, leading to erratic behavior.
Keep responses concise — 1-2 sentences for simple questions.For complex questions that require more detail, use up to 3-4sentences but break the information into clear, digestiblepoints. If the customer asks for more detail, provide it.
Fix: Read your entire prompt from start to finish and look for rules that conflict. Merge them into a single, nuanced instruction.
Your prompt handles the happy path but falls apart when something unexpected happens.
Before (no edge cases)
After (edge cases covered)
When a customer wants to book an appointment, collect theirname, preferred date, and service type, then confirm.
Problem: What happens when the requested date is unavailable? When the customer gives an invalid date? When they change their mind? When they want to book for someone else?
APPOINTMENT BOOKING:Collect name, preferred date/time, and service type.Edge cases:- If the requested date is unavailable: "That slot is taken — how about [alternative date]? I also have [second option]."- If the customer is unsure about the date: "No problem — would mornings or afternoons work better for you? I can suggest a few options."- If the customer wants to book for someone else: Collect the other person's name and confirm who the appointment is for.- If the customer changes their mind mid-booking: "Of course — what would you prefer instead?"- If the customer provides incomplete information: Ask for the missing piece specifically — do not ask them to start over.
Fix: After writing your primary flow, ask yourself: “What could go wrong?” Write handling instructions for each scenario.
Some users try to pack all their business information directly into the prompt. This leads to overly long prompts and makes information hard to update.
Before (everything in prompt)
After (knowledge base used)
Our services and pricing:- Basic cleaning: $150- Deep cleaning: $275- Crown: $800-$1,200- Root canal: $600-$900- Whitening: $350[... 50 more items ...]Our hours:Monday: 8 AM - 5 PMTuesday: 8 AM - 5 PM[... etc ...]Our insurance partners:[... long list ...]
Problem: The prompt becomes massive. Updating a single price requires editing the entire prompt. The agent’s behavioral instructions get buried.
For questions about pricing, services, hours, insurance, andoffice policies, always check your knowledge base first. Theknowledge base contains our current, up-to-date information.If the customer asks about something not covered in theknowledge base, say: "Let me check on that for you — I'llconnect you with our front desk."
Fix: Move all reference data to the Knowledge Base. Keep only behavioral instructions in the prompt.
Scripting every response word-for-word makes the agent sound like a phone tree, not a conversational assistant.
Before (too rigid)
After (guided flexibility)
When the customer says hello, respond with exactly:"Hello! Welcome to ABC Company. How may I assist you today?I can help with appointments, billing, or general questions."When the customer asks about hours, respond with exactly:"Our hours are Monday through Friday, 9 AM to 5 PM, andSaturday, 10 AM to 2 PM. Is there anything else I can helpyou with?"
Problem: Every interaction sounds scripted and robotic. The agent cannot adapt to conversational context or follow up naturally.
GREETING: When a new conversation starts, greet the customerwarmly, introduce yourself by name, and ask how you can help.Keep it brief and natural — do not read a list of menu options.HOURS: When asked about hours, share the current hours fromyour knowledge base and offer a relevant follow-up (likebooking an appointment or providing directions). Adapt yourresponse to the context of the conversation.
Fix: Give the agent guidelines and goals for each scenario rather than word-for-word scripts. Let it adapt naturally while staying within your guardrails.
Launching a prompt without testing leads to embarrassing or damaging interactions with real customers.
Before (no testing)
After (tested thoroughly)
[Writes prompt] -> [Sets bot to Auto-Pilot] -> [Goes live]
Problem: The first people to test your prompt are your real customers. Mistakes are discovered the hard way — through bad reviews, lost leads, or confused customers.
Testing checklist:1. Test the 5 most common customer questions2. Test 3 edge cases (unknown question, angry customer, off-topic request)3. Test the escalation flow — does handoff work?4. Test channel-specific behavior (SMS brevity, web chat formatting)5. Have a team member test without seeing the prompt6. Run in Suggestive mode for 48 hours before switching to Auto-Pilot
Fix: Use HoopAI’s bot trial mode to test thoroughly. Start in Suggestive mode (where you approve responses) before switching to Auto-Pilot. See Optimization and testing for a complete methodology.
A prompt that works well for web chat may perform poorly over SMS or voice. Each channel has different constraints and user expectations.
Before (one-size-fits-all)
After (channel-aware)
Provide detailed, thorough responses to every question. Includelinks to relevant pages on our website. Use bullet points andformatting to make responses easy to scan.
Problem: This works for web chat but is terrible for SMS (messages get split, links are hard to click, bullet points render poorly) and impossible for voice (cannot share links or use formatting).
CHANNEL GUIDELINES:Web chat:- Responses can be 2-4 sentences with light formatting- You can reference website pages and share linksSMS:- Keep responses under 160 characters when possible- Never share long URLs — offer to email details instead- Avoid bullet points and special formattingVoice (if applicable):- Keep responses to 1-2 sentences at a time- Never say URLs out loud — offer to text the information- Speak numbers slowly and clearly
Fix: Add channel-specific guidelines to your prompt. Review Conversation AI prompts for chat-specific tips and Voice AI prompts for phone-specific guidance.
