Conducting Socratic Dialogue

Quick Start

For Socratic dialogue, open with: "Act as a Socratic opponent. I will state a claim. Your job is to ask one pointed question at a time that exposes hidden assumptions, weak evidence, or logical gaps in my argument. Do not agree with me or provide answers — only probe. After each of my responses, either dig deeper into the weakest point or pivot to a new angle."

For adaptive learning, open with: "Explain [topic] to me. I know: [list background, e.g., 'basic algebra, no calculus, work in marketing']. Use metaphors from [domain I know well, e.g., 'cooking' or 'project management']. Check my understanding with a question before moving to the next layer."

Workflow

Socratic Dialogue:

• Step 1: State the claim/argument clearly and completely (no hedging)
• Step 2: AI identifies the single weakest link (assumption, evidence, or logic)
• Step 3: AI asks ONE sharp question targeting that weak link
• Step 4: Respond honestly — defend, revise, or admit the gap
• Step 5: AI either presses further on the same point or moves to the next weakness
• Step 6: Repeat until the argument is either solid or you've mapped all its weak points
• Step 7: Ask AI to summarize which parts of the argument survived scrutiny and which didn't

Adaptive Learning:

• Step 1: Declare current knowledge level and relevant background/analogies domain
• Step 2: AI delivers a first-pass explanation using chosen metaphor
• Step 3: AI asks a check-understanding question (not "does this make sense?" but a concrete test)
• Step 4: Answer the check question
• Step 5: If correct, AI adds one layer of complexity; if incorrect, AI re-explains with a different metaphor
• Step 6: Repeat until target depth is reached
• Step 7: Ask for a final summary in your own words to confirm retention

Examples

Example 1 (Socratic): Input: "Remote work is always more productive than office work because there are fewer distractions." Output: "What counts as a 'distraction' — and is it possible that some in-office distractions (quick clarifying questions, overheard context) actually prevent larger productivity losses later, like rework from miscommunication?"

Example 2 (Socratic): Input: "I revised: remote work is more productive for focused solo tasks, but office work may be better for collaborative tasks." Output: "You've now split the claim by task type — what's your evidence that this split holds across different industries, versus just matching your own personal experience?"

Example 3 (Adaptive Learning): Input: "Explain how neural networks learn. I know basic spreadsheets and nothing about calculus. Use cooking metaphors." Output: "Think of a neural network like a chef adjusting a recipe after each batch of cookies. It tastes the batch (makes a prediction), compares it to the 'ideal cookie' (the correct answer), and tweaks ingredient amounts (weights) slightly based on how far off it was. Over many batches, the recipe converges on something close to ideal. — Quick check: if the cookies come out way too sweet, should the chef make a tiny adjustment to sugar or a drastic one? Why?"

Best Practices

• In Socratic mode, insist the AI ask only ONE question at a time — multiple questions dilute pressure and let you dodge.
• Forbid the AI from offering solutions or validation during Socratic dialogue; its only job is interrogation.
• In adaptive learning, be specific about background ("I've done 2 years of college biology" beats "I'm a beginner").
• Request metaphors from a domain you actually live in daily — generic metaphors (sports, for everyone) are weaker than personalized ones.
• Periodically ask the AI to rate how well your argument/understanding is holding up — explicit scoring prevents false confidence.
• For Socratic dialogue on high-stakes decisions, do a final round asking "steelman the opposite position."

Common Pitfalls

• Letting the AI answer its own questions — defeats the purpose; push back if it does.
• Choosing a Socratic topic you're not actually invested in — the method works best under real argumentative stakes.
• In adaptive learning, accepting a metaphor that "sounds nice" but doesn't map cleanly onto the technical mechanism — ask the AI to explicitly justify the mapping.
• Skipping the check-understanding questions — passive reading of explanations creates illusion of competence without real retention.
• Ending a Socratic session without a summary — without consolidation, you lose track of which objections were actually resolved.