Lesson 16

Why Critical Thinking? How to Ask AI for It

Most people use AI tools like ChatGPT, Claude, or Gemini as high-speed answer machines. They type a question, receive a confident paragraph, and move on. The problem is that these models are designed to be fluent and agreeable, not necessarily correct. They will happily produce a one-sided argument, a glowing recommendation, or a confident summary of a position they have never actually evaluated. If you ask a vague question, you get a vague answer. If you ask a leading question, you get a leading answer. This is not a flaw you can fix by switching tools — it is a property of how these systems work.

The real skill is not getting AI to think for you. It is forcing AI to show its work, consider alternatives, and expose the weaknesses in its own reasoning. This is what prompt engineering for critical thinking means. You are not asking for an answer. You are asking for a structured examination of a question. When you do this well, the AI becomes a sparring partner rather than a yes-man. It will surface risks you had not considered, evidence you did not know existed, and counterarguments that make your own position sharper.

Consider a concrete example. You are a product manager at a mid-sized SaaS company. Your team has proposed migrating your customer support system from Zendesk to a newer, cheaper tool called Intercom. The CFO likes the projected 30% cost saving. The support team likes the modern interface. Everyone is excited. You need to make a recommendation to the executive board next week. If you ask ChatGPT, "Should we switch from Zendesk to Intercom?" you will receive a balanced-sounding but ultimately shallow answer that lists three pros and three cons, then concludes with "it depends on your specific needs." That is not analysis. That is a summary of the first page of a Google search.

Now consider what happens when you ask for critical thinking explicitly. You structure the prompt to demand evidence, trade-offs, and a decision framework. The output changes dramatically — not because the AI suddenly knows more, but because you have forced it to organize what it does know into a form you can actually use.

The Anatomy of a Critical Thinking Prompt

A critical thinking prompt has four components that a simple question lacks. Each one forces the model to behave differently.

• Role and context: Tell the AI who it is and what constraints you operate under. "You are a management consultant with 15 years of experience in B2B SaaS operations. I am a product manager preparing a recommendation for a CFO who is risk-averse and values quantitative evidence."
• Structured analysis request: Ask for specific frameworks — pros and cons, SWOT, cost-benefit, or alternative viewpoints. Do not ask for "thoughts." Ask for a table, a list, or a structured comparison.
• Evidence requirement: Demand that every claim be tied to a reason, a number, or a named source. "For each point, state whether it is a fact, an inference, or an assumption."
• Adversarial instruction: Explicitly ask the AI to argue against the popular position. "Before giving your conclusion, write the strongest possible case for staying with Zendesk, even if you personally think the switch is a good idea."

Here is the full prompt you would use for the Zendesk-to-Intercom decision:

You are a management consultant specializing in B2B SaaS customer operations.
I am a product manager preparing a board recommendation on whether to migrate
our support platform from Zendesk to Intercom.

Context:
- Current Zendesk cost: $2,400/month for 40 agents (Enterprise plan)
- Intercom quoted cost: $1,680/month for the same seats (Pro plan)
- Support team size: 40 agents, 12,000 tickets/month
- Current CSAT: 94% (measured in Zendesk)
- Migration cost estimate: $18,000 in internal labor + 2 weeks of reduced
  response time during transition

Perform the following analysis:
1. List the top 5 pros and top 5 cons of the migration. For each item,
   label it as FACT (verifiable), INFERENCE (based on available data), or
   ASSUMPTION (unverified belief).
2. Run a SWOT analysis (Strengths, Weaknesses, Opportunities, Threats)
   for the migration project itself, not for the tools in isolation.
3. Write the strongest possible case for STAYING with Zendesk. Argue as
   if you are a senior engineer who has seen failed migrations before.
4. Write the strongest possible case for MOVING to Intercom. Argue as if
   you are a growth-focused VP who believes speed matters more than stability.
5. Identify the single most important unknown that would change your
   recommendation, and explain what evidence would resolve it.
6. End with a recommendation: MIGRATE, STAY, or DELAY, with a one-paragraph
   justification based only on the points above.

This prompt works because it does three things at once. First, it gives the model enough context to avoid generic advice — the specific numbers and team size anchor the response. Second, it forces the model to separate facts from inferences from assumptions, which is the core of critical thinking. Third, it demands an adversarial argument, which prevents the model from simply agreeing with whatever position sounds most reasonable at first glance.

Step-by-Step Application

Let me walk you through what happens when you run this prompt in ChatGPT (GPT-4 or newer) or Claude. The exact interface does not matter — the logic is identical.

Step 1: Open a new conversation. Do not add this to an existing chat where you have been discussing the topic casually. A fresh conversation prevents the model from carrying over assumptions from earlier messages. In ChatGPT, click "New chat" in the left sidebar. In Claude, click the "+" icon next to the conversation title.

Step 2: Paste the full prompt above. Do not trim it. Every sentence in the prompt serves a purpose. If you remove the adversarial instruction in point 3, the model will likely produce a weaker counterargument because it has no incentive to argue against the momentum of the conversation.

Step 3: Read the output critically. The first response will be long — often 800 to 1,200 words. Do not skim it. Look specifically at the labels. When the model labels something as ASSUMPTION, ask yourself: "Can I verify this before the board meeting?" When it labels something as FACT, check whether the fact is actually stated in your prompt or whether the model invented it. Models frequently invent plausible-sounding numbers. If the model says "Intercom has a 99.9% uptime SLA" and you did not provide that, it is an inference at best, a hallucination at worst.

Step 4: Run a follow-up challenge. After the first response, type this:

Now challenge your own recommendation. List the three weakest points in
your argument and explain how someone with the opposite view would attack
them. Then tell me what additional data I should gather in the next 48 hours
to strengthen my board presentation.

This second prompt is where the real value appears. The first response gives you structure. The second response gives you depth, because it forces the model to critique its own work — something it will not do unprompted.

Step 5: Extract the decision framework. The final output should give you a clear recommendation with a justification you can defend. If the model says "DELAY," your next action is clear: gather the missing data it identified. If it says "MIGRATE," you now have a list of risks to address in your presentation. If it says "STAY," you have saved your company $18,000 and two weeks of disruption.

Common Mistakes

Expert Tip

Practice Task

Choose a controversial topic you genuinely care about — a policy debate, a technology adoption decision, or a workplace strategy question. Write a single prompt that includes all of the following elements:

• A role and context statement (who you are, what decision you face)
• A request for at least 3 pros and 3 cons, each labeled FACT, INFERENCE, or ASSUMPTION
• A request for the strongest argument against your own current position
• A request for the single unknown that would most change the analysis

Run the prompt in any AI tool you have access to. Then run one follow-up prompt that asks the model to identify the weakest point in its own reasoning. When you are done, check: Did the model label its claims honestly? Did it invent any facts? Did the adversarial argument change your thinking? If you can answer those three questions, you have successfully applied critical thinking to AI — which is more than most users ever do.