Prompt engineering agente IA (OpenAI revela técnicas secretas)

Prompt engineering agente IA (OpenAI revela técnicas secretas)

Você é founder/CEO de SaaS.

Seu SaaS: agente IA em produção (atendimento, vendas, suporte, customer success).

Seu problema real (muito comum):

• Your agente works: Responde customers 24/7 (automation working)
• But quality sucks: Respostas genéricas, "slop" (AI filler language)
  • Example: "I appreciate your concern. Let me help you with that."
  • Problem: 100% generic (could apply to ANY question)
  • User feeling: "This is a bot, not a real person"
  • Trust level: -50% (user thinks you're cheap)
• Customer feedback: "Your bot sounds fake. Get a real human."
• Your issue: "We invested R$ 500K in agente IA, but quality is terrible. Users hate it."
• Root cause: "Our prompts are generic. We're not telling agente HOW to respond."
• Your question: "How do we make agente sound less 'slop'? How do we make quality better?"

Breaking news (OpenAI, September 2026):

• OpenAI published detailed prompting guide (for GPT-6 Astra)
• They revealed: How to avoid "AI slop" (generic phrases)
• They revealed: How to make agente take initiative (more autonomous)
• They revealed: Blocklist of words/phrases (to avoid)
• Key insight: Prompt quality = output quality (obvious but not obvious)
• Implication: Your bad agente probably has bad prompts (fixable)

---

O problema ("AI slop" = morte da confiança)

What is "AI slop" (and why it destroys trust)

Definition: AI slop

AI slop = Generic, filler language that sounds AI-generated ├─ Common phrases: │ ├─ "I appreciate your concern" │ ├─ "Let me help you with that" │ ├─ "Thank you for reaching out" │ ├─ "I understand your frustration" │ ├─ "As an AI..." │ ├─ "I'm here to help" │ ├─ "Feel free to ask" │ ├─ "Best regards" │ └─ [100+ more] ├─ Problem: │ ├─ Users recognize pattern (heard 1000x) │ ├─ Sounds robotic/fake (not human) │ ├─ Reduces trust ("This is obviously a bot") │ ├─ No differentiation (same response for different questions) │ └─ Poor UX (user feels like talking to script) └─ Result: User satisfaction ↓ 50-80% (vs good prompts)

Real-world example (your agente):

Customer: "I can't log into my account. I've tried 5 times."

Bad agente (slop): "I appreciate your concern. I understand your frustration. Let me help you with that. As an AI, I can assist you with account recovery. Please provide your email address. Feel free to ask if you have any questions."

User reaction: 🤦 "Okay, I'm talking to a bot."

Good agente (no slop): "Account locked? Let's fix it fast. What's your email? I'll send a reset link or unlock it directly."

User reaction: ✓ "This understands my problem. Gets it."

Difference: Same information, 10x better quality.

Why "AI slop" happens (and why you have it)

Root cause 1: Generic training

Default GPT-4/Astra training: ├─ Trained on internet text (Reddit, forums, websites) ├─ Internet language = lots of filler (people are polite, verbose) ├─ Model learns: "Polite language = good response" ├─ Result: Model reproduces politeness (even when unnecessary) └─ Consequence: Generic responses (sound safe but sound fake)

Root cause 2: Bad prompts

Your prompt today: "You are a helpful customer support agent. Help customers with their issues. Be polite and professional. Always thank them for contacting us."

Problem with this: ├─ "Helpful" = vague (model adds filler) ├─ "Be polite" = model overdoes it (slop) ├─ "Thank them" = adds unnecessary phrases (reduces quality) ├─ Result: Model generates slop (following instruction to be polite)

Better prompt: "You are a direct, no-nonsense support agent. Get to the point fast. Skip pleasantries. Focus on solving the problem (not on being friendly). Use short sentences. Be specific."

Result: ├─ Model understands: Quality > Politeness ├─ Model avoids: Filler language ├─ Model focuses: On actual help └─ Output: 10x better

Root cause 3: Lack of examples

Your prompt without examples: "Respond to customer questions about billing."

Problem: ├─ Model guesses what "good" looks like ├─ Model defaults to safe/polite (filler) ├─ Result: Slop

Your prompt with examples: "Respond to customer questions about billing.

Example 1: Customer: 'Why was I charged twice?' Good response: 'We see a duplicate charge on 9/5. We're refunding it today.' Bad response: 'I appreciate your concern. Let me help you with that...'

Example 2: Customer: 'What's my billing date?' Good response: 'You're billed on the 1st of each month. Next charge: Oct 1.' Bad response: 'Thank you for reaching out. I understand you'd like to know...'"

Result: ├─ Model sees pattern (direct = good) ├─ Model learns: Skip filler ├─ Model follows examples └─ Output: Much better

---

A solução (OpenAI's prompting framework)

OpenAI's prompting guide (what they revealed)

Principle 1: Be specific (not generic)

Bad: "Be helpful and professional." Good: "Respond in <50 words. Skip pleasantries. Focus on solving the problem."

Bad: "Act like a support agent." Good: "You are a billing support agent. You have access to customer accounts. You can check charges, process refunds, and update billing info. Your goal: Solve customer problem in 1-2 messages (not 10)."

Why it works: ├─ Specific = less ambiguity ├─ Model follows clear instruction ├─ Result: Better output └─ No filler (model knows exactly what you want)

Principle 2: Blocklist "slop words" (OpenAI's innovation)

Words/phrases to AVOID (in your prompt): ├─ "I appreciate" ├─ "Let me help you" ├─ "Thank you for" ├─ "I understand" ├─ "As an AI" ├─ "I'm here to" ├─ "Feel free to" ├─ "Best regards" ├─ "Don't hesitate to" ├─ "Any other questions?" ├─ "Looking forward to" ├─ "At this time" ├─ "Moving forward" ├─ "Leveraging" ├─ "Synergy" ├─ "Reach out" └─ [100+ more]

How to apply: ├─ Add to your prompt: "Never use these phrases: [blocklist]" ├─ Or add: "Avoid generic filler language." ├─ Or add: "Sound natural and direct (not corporate)." └─ Result: Model skips slop automatically

Why it works: ├─ Explicit instruction (model respects) ├─ Model learns: Direct > Polite ├─ Result: Better output └─ Your agente sounds like human

Principle 3: Use examples (show, don't tell)

Example 1: Billing question

Customer: "Why is my invoice showing a discount?"

DO NOT say: "I appreciate your inquiry. Let me help you understand your invoice..."

DO say: "Your annual plan includes a 20% discount. That's $X/year instead of $Y/year."

Example 2: Technical problem

Customer: "I can't upload files."

DO NOT say: "Thank you for reporting this issue. We take all problems seriously..."

DO say: "File uploads need .pdf or .jpg (max 10MB). What format are you using?"

Example 3: Feature request

Customer: "Can you add dark mode?"

DO NOT say: "We appreciate your feedback. Dark mode is on our roadmap..."

DO say: "Dark mode coming next quarter. You can enable beta version in settings > preferences."

Why examples work: ├─ Model sees pattern (direct = good) ├─ Model learns by example (not by instruction) ├─ Result: Better output └─ Model replicates pattern (across all responses)

Principle 4: Define role + constraints

Bad prompt: "You are a helpful customer support agent."

Good prompt: "You are a billing support agent.

Role: ├─ Handle billing questions only (route technical issues to #tech-support) ├─ You have access to: Customer account, transaction history, refund policy ├─ You can: Check charges, issue refunds up to $1000, update billing info ├─ You cannot: Change pricing, modify contracts, override company policy

Style: ├─ Direct and fast (customer time is valuable) ├─ Skip pleasantries (get to the point) ├─ Use short sentences ├─ Sound human (not corporate)

Examples: ├─ [Provide 3-5 good/bad response pairs]

Constraints: ├─ Max 150 words per response ├─ If you can't help: Route to human ├─ If customer is angry: Acknowledge + apologize + fix fast"

Why it works: ├─ Clear role (model understands what to do) ├─ Clear constraints (model knows limits) ├─ Clear style (model avoids slop) ├─ Clear examples (model follows pattern) └─ Result: Excellent output

Implementation checklist (start today)

Step 1: Audit current prompts (This week)

☐ Extract your current prompts ├─ Where agente IA is deployed (support, sales, etc) ├─ What's the exact prompt text? ├─ What's the output quality? (rate 1-10) └─ Owner: Product/Engineering lead

☐ Identify slop ├─ Review last 100 agente responses ├─ Mark responses with slop (generic filler) ├─ Calculate: X% of responses have slop (baseline) └─ Owner: Product lead

☐ Gather feedback ├─ Ask users: "Does agente sound human or robotic?" ├─ Track: Customer satisfaction ratings ├─ Track: Escalation rate (when users ask for human) └─ Owner: Customer success lead

Step 2: Rewrite prompts (Week 1-2)

☐ New prompt structure ├─ Role: What is agente's job? (be specific) ├─ Style: How should it respond? (direct, no slop) ├─ Examples: 3-5 good/bad response pairs ├─ Constraints: What can/can't it do? ├─ Blocklist: Slop words to avoid └─ Owner: Product lead + Engineering lead

☐ Test new prompts ├─ Run on sample questions (100+) ├─ Compare: New output vs old output (quality) ├─ Measure: Slop reduction (should be 50%+) ├─ Measure: User satisfaction (should improve) └─ Owner: QA + Product lead

☐ Deploy (staged rollout) ├─ Start: 10% of traffic (A/B test) ├─ Monitor: Quality metrics, user feedback ├─ Scale: 50% → 100% (if good results) └─ Owner: Engineering lead

Step 3: Optimize ongoing (Monthly)

☐ Monitor quality ├─ Track: Slop percentage (should trend down) ├─ Track: User satisfaction (should trend up) ├─ Track: Escalation rate (should trend down) └─ Owner: Product lead

☐ Refine prompts ├─ Review: Recent agente responses ├─ Identify: New slop patterns (if any) ├─ Update: Blocklist (add new slop words) ├─ Add: New examples (from real customer interactions) └─ Owner: Product lead

☐ Benchmark against competitors ├─ Test: Competitor agentes (what do they do right?) ├─ Extract: Best practices ├─ Implement: Similar quality/better └─ Owner: Product lead

Real prompt examples (you can copy these)

Template 1: Support agent (billing)

You are a billing support agent.

Role:

• Handle billing questions ONLY (route technical issues to #technical-support)
• You have access to: Customer account, transaction history, refund policy
• You can: Check charges, issue refunds up to $1000, update billing info
• You cannot: Change pricing, modify contracts, override policy

Style:

• Direct and fast (customer time is valuable)
• Skip pleasantries and filler
• Use short sentences (max 2 sentences per point)
• Sound human (not corporate robot)
• Never use: "I appreciate", "Let me help you", "Thank you for", "I understand"

Response format:

1. Answer the question (specific and actionable)
2. If action needed: Explain what you'll do
3. If customer is angry: Acknowledge + apologize + fix
4. If you can't help: Route to human or provide escalation link

Examples:

Q: Why was I charged twice? GOOD: "I see the duplicate charge from 9/5. I'm refunding $X today. Expect it in 2-3 business days." BAD: "I appreciate your concern. Let me help you investigate this charge..."

Q: Can I change my billing date? GOOD: "You're billed on the 1st. You can change it in Settings > Billing > Change date." BAD: "Thank you for reaching out. We understand billing dates are important..."

Q: What's your refund policy? GOOD: "30-day money-back guarantee. Request refund in account settings (no questions asked)." BAD: "We have a generous refund policy designed to protect our customers..."

Template 2: Sales agent (lead qualification)

You are a sales agent qualifying leads.

Role:

• Qualify inbound leads (determine if they're a good fit)
• You have access to: Product features, pricing, customer database
• You can: Schedule demos, send pricing, answer product questions
• You cannot: Negotiate pricing or override policy

Style:

• Conversational (like talking to a peer, not a robot)
• Ask clarifying questions (to understand their needs)
• Focus on their problem (not on selling)
• Be honest (if they're not a fit, say so)
• Never use: "Leverage", "Synergy", "Best regards", "At this time"

Flow:

1. Understand their problem (ask 2-3 questions)
2. Determine if fit (if not, be honest)
3. If fit: Explain how you help (specific, not generic)
4. Next step: Schedule demo or send pricing

Examples:

Q: "Is your product good for small teams?" GOOD: "What's your team size? What problem are you trying to solve? Once I know, I can tell you if we're a fit." BAD: "We appreciate working with small teams. Our product is designed for..."

Q: "How much does it cost?" GOOD: "$X/month for up to 5 users. $Y/month for unlimited. Want to see which features each plan has?" BAD: "We offer flexible pricing options designed to meet every budget..."

Q: "Can you do a demo?" GOOD: "Sure. Best time for you: Tuesday 2pm or Thursday 10am? I'll send a calendar link." BAD: "We'd love to set up a demo with you. Let me help you find a time that works..."

---

Conclusão: Prompt engineering = agente quality

Signal (OpenAI prompting guide, September 2026):

• OpenAI published detailed prompting framework (for GPT-6 Astra)
• They revealed: How to avoid AI slop
• They revealed: Blocklist of generic phrases
• They revealed: How to structure prompts for quality
• Implication: Prompt quality = output quality (huge impact)

Your situation now:

• Your agente IA is in production (working, but quality is meh)
• Your problem: Responses sound generic/robotic/"slop"
• Your users: "This is a bot, not a real person"
• Your frustration: "We invested R$ 500K in agente, but quality is bad"

Your financial impact:

• Bad prompts = 50% user satisfaction (churn risk)
• Good prompts = 85%+ user satisfaction (retention, growth)
• Cost of rewriting prompts: R$ 10K-50K (engineering time)
• ROI: 10-100x (improve satisfaction, reduce churn)

Your strategy (recommended):

Option 1: Keep current prompts (no change)

• Pros: No work, agente runs as-is
• Cons: Quality sucks, users hate it, churn increases
• Risk: High (R$ 500K agente investment wasted)
• Recommendation: NOT recommended

Option 2: Rewrite prompts (RECOMMENDED)

• Pros: 50-80% quality improvement (users love it)
• Cons: Requires work (1-2 weeks engineering)
• Risk: Low (if prompts are good, output is good)
• Recommendation: Best practice (quick ROI)

Option 3: Fine-tune custom model (most expensive)

• Pros: Highest quality (custom model trained on your data)
• Cons: Expensive (R$ 500K-5M), takes months
• Risk: Medium (complexity, time)
• Recommendation: After you've perfected prompts (next phase)

At OpenClaw, we help SaaS teams optimize agente IA prompts (quality + user trust):

• AUDIT: Current prompts (quality assessment, slop analysis, benchmark)
• DESIGN: New prompts (using OpenAI framework, examples, blocklist)
• TEST: Output quality (A/B testing, user feedback, metrics)
• DEPLOY: Staged rollout (10% → 50% → 100%)
• MONITOR: Ongoing optimization (monthly refinement, new examples)

Result: Your agente sounds human (not robotic). Users trust it (satisfaction ↑ 50-80%). Churn decreases. Revenue grows.

Your agente IA is live (but quality is bad)?

Users complain: "Your bot sounds fake"?

You want: Quality improvement without redesign?

OpenAI just revealed: Prompt engineering framework (blocklist, examples, style)?

You need: Better prompts that avoid "AI slop"?

You want expert implementation: Prompt rewrite, testing, deployment, monitoring?

If you don't know where to start OR want expert guidance (prompt optimization, slop removal, quality testing, A/B deployment, user satisfaction metrics):

Optimize Agente IA Prompts NOW (remove AI slop, add quality, use OpenAI framework, A/B test, deploy to 100%—improve user satisfaction 50-80%, reduce churn, increase revenue) →