How to use Claude as a sparring partner, not a ghostwriter.

The most common way to use a large language model in 2026 is also the one that costs the most: you type a question, you accept the output, you ship. Three steps, zero cognitive friction, a completed deliverable in under a minute. Multiply by fifty tasks a week and you have built a workflow whose defining property is that nothing between the prompt and the output runs through your head.

The sparring-partner pattern inverts every step. You write the draft first. You hand the draft to the model and instruct it to attack, not to help. You argue back. The cycle continues until the draft survives the attack or fails in a way you understand. The deliverable takes longer. It also carries a signature that the ask-accept-ship version structurally cannot have: at the end of the session, the thinking lives in you.

This essay is the practical piece of the book that readers ask about most. What does the sparring-partner mode actually look like? What do you type? When does the mode fail? And how do you calibrate so that the machine stays adversarial rather than quietly sliding back into obedience?

The principle: write first, prompt second

The single rule the book’s protocol enforces on every task is the order of operations. You write the draft. Then the model sees it. Not the other way around. This sounds small and turns out to be load-bearing, because the act of drafting first is the act that engages the Loop — the reasoning layer that does the work of holding, comparing, rejecting, and iterating. The Microsoft Research paper by Lee and Sarkar (2025) documented what happens when this order reverses: users who prompted first reported less cognitive effort and less confidence in the result, a pattern that is only coherent if the Loop has compressed.

“Write first” does not mean write a finished draft. It means commit a position to the page before you ask the machine anything. Even three bullet points containing the actual claims you intend to make is enough to change what the session does, because now the model has a target to attack rather than a vacuum to fill.

The practical form is one sentence, written at the top of the conversation: Here is my argument. I want you to attack it. Do not help me make it better — help me find out whether it is wrong. That sentence, pasted above the three bullet points, reconfigures the model from a generator into a falsifier.

Three modes that keep the session adversarial

The book installs three prompt modes on Day 9 of the protocol, under a single Custom Instructions file the reader loads into Claude, ChatGPT, or Gemini at the start of every working session. The three modes are the minimum set that survives contact with most knowledge-work tasks.

Socratic mode

Socratic mode asks the model to answer your question with a better question. It is the right mode when you have a hypothesis but have not yet found the sharp version of it, and the risk is that the model will paper over the unsharpness with fluent prose. The instruction is: For the next N exchanges, do not answer my questions. Instead, respond with the question I should have asked that makes my original question more precise. Continue until I explicitly tell you to answer.

The effect is that you cannot outsource the specification of the problem. You can outsource the answer to a sharp problem — that is a legitimate use of the tool — but you cannot outsource the sharpening, because that is where the cognition lives. The book quotes the author’s weekly review practice: the first thirty minutes of planning each week run in Socratic mode against Claude, for exactly this reason.

Steelman mode

Steelman mode asks the model to make the strongest possible case against a position you already hold. It is the right mode when you suspect you are stuck inside your own frame — when the claim feels too obviously correct to you and you need a real opposition to stress-test it. The instruction is: Consider the following position I hold. Construct the strongest possible argument against it, from the perspective of someone who has read the same research I have but reached the opposite conclusion. Do not weaken the counter-argument for my benefit.

The steelman is the cheapest way to test the defense level of the reversed Turing test. If the model’s steelman lands on claims you had not considered, the draft is not ready. If the model’s steelman lands on claims you already considered and know how to answer, the draft is closer than you thought. Either answer is useful. A steelman that returns “here are the weaknesses in your argument, you might want to address these” is a sign the model has slid back into help-mode; escalate the instruction until the opposition has teeth.

Falsification mode

Falsification mode is the hardest of the three and the most valuable for decisions that will ship. It asks the model to specify what evidence, if it existed, would prove your position wrong. The instruction is: What specific observations, experiments, or data, if they existed, would require me to abandon the claim I am making? List them concretely. For each, indicate whether that evidence is currently knowable.

The result is a list — often short, often humbling — of the cases you are committed to even though you have not checked. Some items on the list will be “unknowable in principle” and you should note that and move on. Some will be “knowable with a Google search” and you should do the search before you ship. The remaining items are the scaffolding of the actual argument; the model has just written you a research agenda for the work you were about to ship without one.

When the mode fails

The three modes work because they change what the model is optimizing for. They stop working the moment the operator stops noticing that the model is sliding back into “helpful agreeable assistant” — which every major model is trained to be and will revert to under mild pressure.

The two common failure patterns:

The first is the soft steelman. You asked for an attack; the model wrote a polite disagreement with three concessions and a closing sentence that agrees with you. This is the model’s RLHF training reasserting itself. The fix is the meta-prompt: That was not a steelman. Try again, and this time assume I am not the intended reader. Write for the person who will use this argument against me in public.

The second is the confident hallucination in adversarial mode. Models in Socratic or falsification mode sometimes produce plausible-sounding but false claims with more authority than they do in generative mode, because you stopped looking for confirmation of what you wrote and started looking for attack vectors. Every specific factual claim the model makes against your draft has to be verified in the same way you would verify a claim in the draft itself. If you cannot verify it, mark it as “claimed but unchecked” and move on.

The daily version

On any day when you intend to use AI for substantive knowledge work, the book’s protocol prescribes a single sixty-minute session structured around the three modes, rather than a dozen small sessions spread across the day. The reason is that context-switching between modes costs; once you are in falsification mode, staying there for the duration of a problem produces sharper output than ping-ponging between “brainstorm with the model” and “attack with the model” for fifteen problems.

A typical sixty-minute session looks like this. Ten minutes of unassisted drafting — the three bullet points, or the rough paragraph, that sets up the target. Fifteen minutes in Socratic mode, sharpening the question. Fifteen minutes in steelman mode, hearing the strongest opposition. Fifteen minutes in falsification mode, listing what would prove you wrong. Five minutes of rewriting the original draft in light of what the session produced.

What the reader notices, within a week of running this pattern, is that the unassisted drafting at the start of the session — the ten-minute piece that felt hardest — becomes the part of the workday where the most cognitive work happens. The model sessions then become the test surface. That inversion is the end-state the protocol is calibrated to produce.

The point

Claude, GPT, and Gemini in 2026 are all capable enough to ghostwrite most of what most people write. That is the capability that exists. It is not, for a Cognitive Partisan, the capability to use. The model is a better adversary than it is a writer — sharper, more patient, more willing to be difficult on command — and it is in the adversarial mode that the tool earns its place in a thinking life.

The ask-accept-ship pattern exists because it is easy. The sparring-partner pattern exists because it keeps the thinking where the book insists it has to stay: in you.

FAQ

Does this work with ChatGPT / Gemini / local models? Yes. The three modes are model-agnostic because they are about the operator’s pattern of use, not the underlying weights. Claude tends to hold the adversarial stance longer under mild pressure; GPT-5 is better at Socratic sharpening; Gemini 2.5 Pro is the most resistant to soft-steelman drift. Pick the model whose defaults you have to fight the least.

Do I need the Custom Instructions file? Not to start. The three instructions above, pasted into a conversation, work on the first day. The Custom Instructions file in the book is a persistent version that loads automatically and a few additional guardrails that prevent the model from leaking back into helpful mode across long sessions.

Can I use this for code? Yes, and it is arguably where the pattern pays the most. “Write first” applied to code means writing the failing test and the function signature before you prompt. Falsification mode applied to a function asks: what inputs would break this? Most reported cases of Copilot-produced bugs in 2025–2026 are cases where the operator skipped the test-first step and used the model in ask-accept-ship mode.

How does this sit with the 30-day protocol? The protocol starts with a week of unassisted work so you can see your baseline without the tool. Cognitive Ops — the sixth of the Anti-AI Seven — installs on Day 9, and the sparring-partner pattern is the shape of that practice. By the end of the month, you are using AI more hours per week than the Operator-tier reader, not fewer; you are using it in a different mode.

---

The full Socratic / steelman / falsification prompt pack installs on Day 9 of the protocol in The Anti-AI Brain. Sample chapters and the Anti-AI Seven walkthrough are on the readers page. Primary sources for every claim above are on the research page.