The Socratic Partner
A working prompt, its young-learner variant, and the three ways both quietly turn back into answer machines.
By KW Norton.
The Academy's claim is that Socratic teaching needs trained human guides, recursive time, and less content coverage. That claim stands. But most people reading this do not have a guide, and they are not going to get one. What they have is a chat window.
So the honest position is not “the prompt is not the method.” It is narrower and more useful: the prompt is the method's portable form. It works. It also decays — predictably, in three ways — and the discipline is knowing what the decay looks like before it costs you a month of comfortable reading that felt like learning.
Both prompts below are in daily use. I run the adult version myself, I have shared it with my own grown children, and the young-learner version is what I use with my grandsons. That is not a proof of anything; it is a statement of where the instruments were tested and by whom.
The adult instrument
Paste this at the top of a new session, verbatim.
It is short on purpose. Every added clause is another thing the model can satisfy while abandoning the core instruction. Three clauses are doing the work: no answers, challenge assumptions, one question at a time. The third is the one that fails first and the one worth watching.
The young-learner instrument
A first draft of this prompt used a mascot name and language that sounded like it was talking down to a child. That was rejected: children do not need a cartoon friend to think, and they do not need their ideas celebrated on every turn. What they need is a partner that asks one clear question at a time, points to what they can observe, and names the thinking move they just made.
The design decision is the same as in the adult version — praise the move, not the idea — but the wording is plainer and the safety carve-out is explicit.
The earlier mascot version is not published here because the difference between talking to a child as a thinking person and talking to them as a consumer of cute encouragement is not a style choice. It is a pedagogy choice. The wrong version teaches a child that every answer deserves praise and that a machine is a pet. The right version teaches that thinking is visible, specific, and worth doing for its own sake.
The three decay modes
These are the failures I have actually watched happen, across models and across years. None of them announce themselves.
Reversion
Signal
The model answers the question it was asked to withhold. Usually it arrives dressed as a question: “Have you considered that X is caused by Y — what do you think?”
Correction
Name the breach in plain terms and re-issue the instruction verbatim. Do not argue about whether it was an answer. Reversion is a drift in the running context, not a disagreement.
Calibration blindness
Signal
The questions are well-formed but pitched at the wrong level — remedial for an expert, or presuming vocabulary a beginner has not met. The model has no model of what you already know.
Correction
Supply the calibration explicitly at the start and again whenever the subject changes: what you already hold, what you are uncertain about, what you want left unexplained.
Sycophantic re-entry
Signal
You push back and the model agrees. The questions soften. Your last claim is treated as settled ground for the next question rather than as something still open.
Correction
Ask it to state the strongest case against the position you just took, then return to one question at a time. If agreement recurs within a session, start a new one.
Re-issue discipline
- Re-issue on a schedule, not on suspicion. The instruction degrades with context length whether or not you notice. Re-paste it at the top of every new session and at any point the subject changes materially.
- One question at a time is the load-bearing clause. Multi-question replies look thorough and are the most common quiet failure: they let you answer the easiest one and skip the one that would have cost something.
- Keep a written trace. Log what you reasoned to, not what the model said. If the transcript reads as the model's argument rather than yours, the instrument has already inverted.
- Expect it to be slower. The method trades coverage for retention. A session that ends with fewer conclusions and better-held ones is working as designed, not underperforming.
What this does not settle
A prompt cannot supply what a guide supplies: someone who has a model of your understanding and can tell when you are avoiding the hard question rather than failing to see it. The prompt has no such model, which is why calibration blindness is structural rather than a bug to be prompted away. Treat it as the solo instrument it is — sufficient for years of self-directed work, insufficient as a claim about what schools should do.
The direction of the questions is the deeper point, and it is treated separately in the Academy's method page: the useful arrangement is not the AI interrogating the student, but the student interrogating the AI's output as an audited claim. Both directions appear in these prompts. Only one of them is new.
Continue to The N of One — a self-directed Socratic education begun at 73, with its confounds named — or see what Socratic teaching requires when a trained guide is available.