Meta-prompting, self-consistency, prompt chaining, and tree-of-thought.
Once the core techniques are comfortable, these advanced methods unlock more sophisticated uses. They're worth the effort for complex or high-stakes tasks.
What it is: You ask the AI to help you write a better prompt for the actual task you want to accomplish.
When to use it: When you have a complex task but aren't sure how to structure the prompt. Also useful when earlier prompts have produced disappointing results.
Example: > "I want to ask you to help me analyze customer churn data and identify the most likely root causes. Before I do, help me write the best possible prompt for this task. Ask me any clarifying questions you need."
The AI may ask about data format, what decisions this analysis will inform, what you've already tried, and what success looks like. Then it helps you formulate a prompt that's much more likely to produce useful output.
Works best with: Claude and GPT-4 class models that handle meta-reasoning well.
What it is: You ask for multiple independent answers to the same question, then identify which is most consistent, most common, or most defensible.
When to use it: Complex reasoning, math, analysis, or situations where you want to pressure-test an answer. Also useful for creative tasks where you want genuine variety.
Example: > "Give me three independent analyses of the following business decision, each starting from scratch. After all three, tell me which conclusion was most consistent across all three analyses and why.
Decision: [describe decision]"
Self-consistency is particularly valuable for reasoning tasks where the AI might take a wrong turn early. Running multiple paths and comparing them surfaces errors that a single answer would hide.
Note: This uses more tokens and costs more if you're on a metered API. Worth it for important decisions.
What it is: You break a complex task into a sequence of steps, where the output of each prompt becomes the input to the next.
When to use it: Anytime a task is too complex to do well in a single prompt. Research → outline → draft → edit is a classic chain. Analysis → synthesis → recommendation is another.
Example chain for a competitive analysis: 1. "List the top 5 competitors to [company] in [market]. For each, give: name, core product, primary target customer, and pricing model." 2. "Based on the competitive landscape above, identify the three most significant gaps or unmet needs in this market." 3. "Given these gaps, suggest three product differentiation strategies [company] could pursue. Rank them by feasibility and potential impact."
Each step produces better output because it's focused on one thing. The chain builds context incrementally rather than trying to solve everything at once.
What it is: You ask the AI to explore multiple different approaches or reasoning paths before committing to an answer.
When to use it: Complex problem-solving, strategic decisions, situations where the right answer isn't obvious and you want the model to consider alternatives before concluding.
Example: > "Before answering, consider three different ways to approach this problem, briefly noting the pros and cons of each. Then choose the best approach and explain your reasoning.
Problem: We need to reduce customer support ticket volume by 30% in 90 days without increasing headcount."
Tree of thought slows the model down from jumping to an answer and encourages it to evaluate alternatives. It reliably improves quality on complex reasoning tasks.
Works especially well with: Claude and GPT-4o for structured reasoning tasks.
What it is: A framework where the AI alternates between reasoning about what to do next and then acting (taking a step). Each action produces new information that informs the next reasoning step.
When to use it: Multi-step research tasks, debugging processes, any task where each step might reveal information that changes the approach.
Prompt structure: > "Solve this step by step using the following pattern: > Thought: [what you're thinking about the current situation] > Action: [what you do next] > Observation: [what you learn from that action] > Repeat until you reach a conclusion.
Task: [your task]"
ReAct is particularly valuable when used with AI agents that have access to tools (web search, code execution). It structures the AI's decision-making process in a way that's more transparent and easier to verify.
What it is: You provide the AI with a set of explicit principles or constraints that should guide all of its responses on a topic.
When to use it: When you need consistent behavior across complex or nuanced situations. Legal review, content moderation, communication guidelines, editorial standards.
Example: > "You are reviewing customer-facing communications for our company. Apply the following principles to every review: > 1. Never make claims we can't verify or substantiate > 2. Avoid language that could be interpreted as a guarantee unless we explicitly offer one > 3. Flag any statement that implies comparison to competitors without citing data > 4. Maintain a confident but not arrogant tone > > Review the following: [content]"
The principles act as a consistent standard the model applies throughout. This is closer to how AI systems are built for high-stakes use cases — it's essentially a miniature system prompt within a conversation.
These techniques are worth learning but not worth applying to every task. Zero-shot prompting works fine for the vast majority of interactions. Pull out the advanced techniques when: - The task is complex and the stakes are high - Earlier attempts have produced unsatisfying results - You need consistent behavior across many iterations - You're building something where prompt quality compounds — an agent, a workflow, a product
The best prompters know which tool fits which job.
Have a follow-up question about this topic?
Ask AI