Advanced Prompting Techniques (For When Basic Prompts Aren't Cutting It)
You've mastered the basics. Your prompts are specific, structured, and don't sound like a corporate robot wrote them. Great.
Now let's get weird.
These are the advanced techniques I use when I need the AI to do something it's not naturally good at — or when I want output that doesn't look like every other AI-generated thing on the internet.
1. Chain-of-Thought Prompting
What it is: Making the AI "show its work" before giving an answer.
When to use it: Complex reasoning, analysis, problem-solving.
Basic prompt: "Should I launch this product now or wait 3 months?"
Chain-of-thought prompt:
"I'm deciding whether to launch a product now or wait 3 months. Before answering, think through:
1. What are the pros/cons of launching now?
2. What are the risks of waiting?
3. What assumptions am I making?
4. What would change my decision?
Then give your recommendation with reasoning."
Why it works: Forces the AI to think step-by-step instead of pattern-matching to a generic answer. You get better reasoning and can spot flawed logic.
When AI jumps straight to conclusions, you don't know how it got there. Making it show its work lets you evaluate the thinking, not just the answer.
2. Role + Constraints + Anti-Patterns
What it is: Defining what you DON'T want as much as what you DO want.
When to use it: When you keep getting output that's technically correct but stylistically wrong.
Example:
"You are a senior product manager who's launched 20+ products. Write a product requirements doc for a mobile app feature. Include: user stories, acceptance criteria, edge cases. Do NOT: use corporate jargon, write vague requirements like 'intuitive UI', or skip the 'why' behind each requirement. Tone: clear, opinionated, slightly impatient with ambiguity."
Why it works: The "do NOT" section acts like guard rails. Prevents the AI from falling into its default patterns.
Think of it like training a dog. "Sit" is one instruction. "Sit, don't jump, don't bark" is more specific and gets better results.
3. Few-Shot Prompting (Give Examples)
What it is: Showing the AI 2-3 examples of what you want before asking it to generate.
When to use it: Specific formats, unique styles, anything where "explain it in words" is hard.
Example:
"Write 5 email subject lines for a SaaS product announcement. Here are 3 examples of the style I want:
• 'We built the thing you asked for (finally)'
• 'This feature broke our beta testers' brains'
• 'Why we killed our most-loved feature'
Notice: Conversational, intriguing, not clickbait-y. Now write 5 more in this style for [your product]."
Why it works: Examples are worth a thousand words of style instructions. AI is great at pattern-matching when you give it the pattern.
4. Perspective Shifting
What it is: Asking the AI to write from a specific, unusual perspective.
When to use it: When you want fresh angles or need to challenge assumptions.
Example:
"Write a critique of our product roadmap from the perspective of:
1. A customer who loves us but is frustrated
2. A competitor who's trying to beat us
3. Our future selves in 2 years looking back
For each perspective, identify what we're missing or getting wrong."
Why it works: Different perspectives surface different insights. The "future self" one is especially good for catching short-term thinking.
This is basically Red Team thinking. You're stress-testing ideas by attacking them from multiple angles.
5. Constraint Stacking
What it is: Adding multiple creative constraints to force novelty.
When to use it: When normal prompts give you boring, predictable output.
Example:
"Write a blog post about productivity. Constraints:
• Must use a food metaphor throughout
• No bullet points or listicles
• Written as a letter to your past self
• Include one counterintuitive claim
• 400 words max"
Why it works: Constraints breed creativity. The AI can't fall back on templates when you've made the template impossible.
It's the Twitter/haiku principle. When you have infinite freedom, you get generic. When you have strict limits, you get creative.
6. The "Devil's Advocate" Prompt
What it is: Asking the AI to argue against your position or poke holes in your logic.
When to use it: Before making big decisions, launching things, or when you need a reality check.
Example:
"I'm planning to [your plan]. Play devil's advocate. What are the strongest arguments against this? What am I not seeing? What assumptions might be wrong? Be brutally honest — I want the criticism, not reassurance."
Why it works: Forces you to confront weak points before they become problems. Also prevents groupthink when everyone around you is nodding along.
The key phrase: "brutally honest." Without that, AI defaults to polite and unhelpful.
7. Iterative Refinement (The "Sharpen This" Loop)
What it is: Getting output, then asking the AI to critique and improve its own work.
When to use it: When the first draft is close but not quite there.
Example:
[AI generates something]
"Now critique this. What's weak? What's unclear? What could be cut? Then rewrite it addressing those issues."
[AI improves it]
"One more pass. Make it 20% more concise without losing key points."
Why it works: The AI is surprisingly good at self-editing when prompted. Gets you to a polished final draft faster than manually tweaking.
8. The "Explain Like I'm Five, Then Like I'm a PhD" Technique
What it is: Getting both the simplest and most complex explanation of something.
When to use it: Understanding new topics, preparing to teach/present to different audiences.
Example:
"Explain [complex topic]:
1. Like I'm 5 years old (simple metaphor, no jargon)
2. Like I'm a PhD in this field (technical depth, nuance, current debates)"
Why it works: The ELI5 version tests if the AI (and you) actually understand it. The PhD version surfaces the complexity and edge cases. Together, you get full spectrum understanding.
9. Output Format Control
What it is: Specifying exact output structure, not just content.
When to use it: When you need to copy-paste directly into another tool or format.
Example:
"Generate 10 tweet ideas about [topic]. Format as a JSON array with keys: 'tweet_text', 'hook_type', 'target_audience'. No additional commentary."
Or:
"Create a comparison table: React vs Vue vs Angular. Columns: Framework, Learning Curve, Ecosystem, Best For. Format as Markdown table."
Why it works: Saves you from reformatting. Gets you machine-readable output when you need it.
10. The "Simulate a Conversation" Technique
What it is: Having the AI role-play both sides of a dialogue to explore ideas.
When to use it: Preparing for tough conversations, exploring different argumentative angles.
Example:
"Simulate a conversation between:
• A founder who wants to pivot the product
• A skeptical investor who thinks it's too risky
Make both sides smart and well-argued. 3-4 exchanges each."
Why it works: Surfaces arguments you might not think of. Helps you stress-test ideas through dialectic.
11. The "What Would X Do?" Prompt
What it is: Channeling a specific person's thinking style.
Example:
"How would [specific person] approach this problem? Think about their framework, priorities, and typical advice. Then give your answer in their style."
Try: "How would Tim Ferriss optimize this process?" or "How would a Stoic philosopher think about this setback?"
Why it works: Imports a specific mental model instead of generic advice.
12. The "Make It Worse" Technique
What it is: Asking for the opposite of what you want to identify failure modes.
Example:
"I'm writing a welcome email for new users. Instead of writing a good one, write the WORST possible welcome email. What would make users immediately unsubscribe?"
Then you know exactly what to avoid.
Why it works: Sometimes it's easier to identify what NOT to do than what TO do. Inversion is a powerful thinking tool.
When to Use Advanced Techniques
Don't use these all the time. They're overkill for simple tasks.
Use them when:
- Standard prompts give generic output
- You need the AI to think, not just regurgitate
- The stakes are high (decisions, launches, important communication)
- You want output that doesn't look AI-generated
For "write a tweet" or "summarize this," basic prompts are fine. For "help me think through this complex decision," bring out the advanced toolkit.
Combining Techniques
The real power move? Stack these together.
Example:
"You are a product strategist with 15 years experience. Use chain-of-thought reasoning to analyze whether we should build feature X. Consider the perspectives of: our power users, casual users, and our future selves in 1 year. Present pros/cons in a table. Then play devil's advocate and argue against your own recommendation. Be brutally honest."
That's role + chain-of-thought + perspective shifting + devil's advocate + format control all in one prompt.
And it'll give you better analysis than 90% of consultants.
The Meta-Lesson
Advanced prompting is about controlling the AI's reasoning process, not just its output.
You're not just saying "write this" — you're saying "think like this, then write."
You're directing not just what it produces, but how it thinks about the problem.
Master these, and you'll get AI output that makes people ask, "Wait, did you actually write this yourself?"
Which, honestly, is the highest compliment.