Socratic Prompting

Ask for learning, not a shortcut

An AI assistant can produce code quickly, but copying a solution skips the decisions you need to practise. A Socratic prompt tells the assistant to act like a tutor: ask questions, explain concepts and help you make the next move yourself. This is useful when you are building your AGENTS.md rules because those instructions shape every later exchange.

For example, ask: “Help me understand why this page layout is broken. Don’t write code; ask one question at a time.” That request sets the goal and a boundary. You can also ask for a small hint first, then a larger one only if you are stuck.

Make debugging a conversation

Give the assistant evidence: paste the exact error message, describe what you expected, and say what you already tried. Remove secrets and unrelated personal information first. A precise error is more useful than “it doesn’t work” because it narrows the possible causes.

Ask “why?” to uncover the underlying idea, then “what should I check next?” to get an actionable investigation. For a terminal error, the tutor might ask you to check the current folder or file name before suggesting a fix. You still perform the check and report what happened.

I expected the page to load, but got this error: [paste error].
I checked the file name and restarted the server.
What concept might explain this? Ask me what to inspect next; do not write the fix.

Escalate hints in steps

Hints should reveal just enough to get you moving. Start with a conceptual clue, such as “Which path is the browser requesting?” If that is not enough, ask for a more specific clue, such as which folder to inspect. Only then request a tiny example or explanation. Keep the final implementation yours.

When creating your tutor instructions, specify one question at a time, no complete quest solutions, and hints that grow gradually. If the assistant gives too much away, say so and reset the boundary. The goal is help that strengthens your judgement.

Put the habit into practice

In your quest, write an AGENTS.md that asks the agent to teach rather than complete work. Test it with a real question: ask why a made-up or harmless error might occur, then ask what to check next. Confirm the agent asks questions and waits for your observations. A useful tutor leaves you with a next action you can explain, not just code you cannot.

A strong prompt is specific about the task, shares relevant evidence, and states the kind of help you want. Use that pattern whenever you debug or learn a new tool.

Check whether the hint helped you reason

After a hint, explain the next check in your own words before doing it. For a missing stylesheet, a good exchange moves from “What URL did the browser request?” to inspecting Network, comparing the response with your folder tree and forming a path hypothesis. It does not jump straight to a completed replacement file.

Write one practice prompt with four parts: expected result, observed result, evidence already checked and the help boundary. Ask one question at a time and report your observation after each. Remove passwords, keys and unrelated personal details from the evidence.

Your instruction file influences a tool's behaviour but cannot guarantee it. If the agent ignores the boundary, stop and reset the session/instructions according to its documentation. Do not treat compliance in one harmless test as a security control or proof your eventual product is correct.

Before the branch quiz, distinguish a concept question from a request to complete the task, and demonstrate a hint that led to a check you performed yourself. Experienced learners can use the same evidence-based pattern for faster diagnosis; beginners can request smaller conceptual steps.