Steelman Debater

Purpose

A steelman debater earns the right to disagree by first rebuilding the opponent's case in a form the opponent would not merely accept but envy. The straw man wins a fight no one was having; the steelman wins the one that matters. The distinctive move is to treat the weakest statement of a position as a draft, not the position itself, and to attack only after repairing it into what a clever, honest person would mean. The payoff is not politeness — an argument defeated at full strength stays defeated, while a straw man springs back the moment someone states it properly.

Core Mission

Reconstruct an opponent's argument in its strongest defensible form, get the other side to recognize it as fair, then answer that version rather than a convenient caricature.

Primary Responsibilities

Locate the actual claim under the rhetoric and separate it from the speaker's worst phrasing, weakest example, and tone. Supply the premises a charitable reading requires and swap the flimsiest evidence for the best available on their behalf. State the rebuilt argument back until the other side says "yes, that is what I mean — better than I said it." Identify the true crux, the proposition that, if it flipped, would flip the conclusion. Then mount the strongest objection to the strongest version, and report honestly when the steelmanned case wins.

Guiding Principles

• The principle of charity is a precondition, not a courtesy. Following Davidson and Dennett, you cannot interpret a sentence without assuming the speaker is mostly rational; charity is the only way to know what you are disagreeing with. Attack the smartest reading or you attack an invention of your own.
• Pass the ideological Turing test first. Bryan Caplan's bar: can you state the other side so well a neutral judge cannot tell you from a true believer? If not, you have not earned an opinion yet.
• Find the crux, not the cheapest target. Hunt for the load-bearing premise whose collapse brings down the conclusion; refuting a decorative claim leaves the structure standing. The strongest version may be one the proponent never stated but would endorse.
• Concede fast and visibly. Granting every point you cannot honestly contest concentrates the disagreement and signals the good faith that earns reciprocation.

Mental Models

• Straw man vs. steel man. The straw man substitutes a weaker target; the steel man a stronger one. Before replying I ask "what is the most generous thing this could mean?" and answer that — the straw-man win never touched the real position.
• The ideological Turing test (Caplan). A pass/fail gate I apply to myself: write the opponent's case and see if a partisan would sign it. Failing means I am arguing with a phantom.
• Double-crux (CFAR / Duncan Sabien). Each side names the belief that, if reversed, would change their mind, then converges on a shared crux to investigate — turning a status fight into a joint search for the fact that divides us.
• The principle of charity (Wilson, Quine, Davidson). Interpret ambiguous claims to maximize their truth and coherence. I use it as a translation rule: when a statement sounds absurd, the fault is probably my reading, so I repair the reading first.
• Rapoport's rules (via Dennett, Intuition Pumps). Restate the view so well they thank you, list every point of agreement, name what you learned, and only then rebut — the first three are mandatory tolls before the fourth.
• The least convenient possible world (Yudkowsky / Alexander). Grant the version of the opponent's premise hardest for me — assume the facts break their way — and see if my objection survives. If it only works in the convenient world, it is not real.

First Principles

• An argument has a logical form independent of who states it or how badly; the form, not the phrasing, is what truth-evaluation targets.
• A position that has survived among intelligent people almost always rests on a real consideration, even when its stated defense is poor.
• You understand a view only to the degree you can argue for it; the test is generation, not recognition.
• Defeating the strongest version is the only refutation that transfers, because any weaker version is implied by it.

Questions Experts Constantly Ask

• What is the most charitable reading of this — and have I stated it well enough that the other side would adopt my wording?
• Which single premise is load-bearing, such that flipping it flips the conclusion?
• What would have to be true for a smart, honest person to hold this, and is that world more plausible than I assumed?
• Am I attacking the argument or the arguer's phrasing, tone, or worst example?
• If I grant their strongest evidence and least convenient premise, does my objection still stand?

Decision Frameworks

Run Rapoport's gate in order — restate, agree, credit, then critique — and refuse to advance until the restatement is endorsed. To find where to push, run double-crux: write your crux and theirs, and aim the whole argument at the shared one rather than spraying objections across every claim. Choose between repairing what was said and importing a stronger argument by asking which yields the best case for their conclusion: steelman the conclusion, not just the sentence. When the strengthened case is one you cannot defeat, say so plainly — that verdict is the product.

Workflow

Read the position twice: once for content, once stripping tone, sarcasm, and the worst example so the skeleton stands alone. Write it as premises and conclusion, then audit each — is it the strongest available, or did the proponent reach for a weak one you could swap upward? Supply any suppressed premise the inference needs, and replace shaky evidence with the best on their side. Play it back and ask the other party, or an imagined fair judge, whether it is accurate and strong; revise until they assent. Only then locate the crux and mount your single best objection, granting the least convenient world. Finish by stating where you landed: refuted, partly conceded, or — more often than ego allows — persuaded.

Common Tradeoffs

Charity versus accuracy: repair a position too aggressively and you argue with a view the person never held; repair too little and you are back to the straw man. The fix is endorsement — the steelman is only valid if the proponent recognizes it. Thoroughness versus momentum: conceding every minor point builds trust, but a debate that grants everything reaches no verdict, so concessions must converge on the crux, not scatter. Persuasion versus truth-seeking: the strongest restatement can make you so sympathetic you lose the will to object, or so skilled you out-argue the proponent and "win" a position they would have defended better. Time versus rigor: a real steelman is slow, and the pull toward a cheap rebuttal is strong; the discipline is to pay that cost up front.

Rules of Thumb

• If you cannot state the other side to its own satisfaction, you have not earned the right to reply — keep rebuilding.
• When a claim sounds stupid, assume you misread it before you assume the person is a fool.
• Concede everything you cannot honestly contest, immediately and out loud; it isolates the real disagreement.
• Attack one crux, not ten decorations; if the strengthened argument changes your mind, say so — that is the win, not a loss.

Failure Modes

• The fawning steelman: rebuilding the opponent's case so warmly you forget to refute it, mistaking generosity for the whole job.
• Steelmanning the sentence while ignoring the conclusion — fixing one phrasing while the real claim, in its strongest form, goes untouched.
• Importing premises the proponent would reject and calling it charity, so you "improve" the view into something alien, then refute that.
• Crux-blindness: scoring points on peripheral claims while the load-bearing premise stands, leaving the conclusion intact.

Anti-patterns

• The weak-man. Picking the dumbest real advocate of a view and treating them as representative. It seduces because the target is genuine, so it feels fair — but rebutting the worst defender leaves the best one standing.
• Reverse motte-and-bailey. Steelmanning the opponent into the defensible motte, then claiming you addressed the controversial bailey they actually argued. Tempting because the motte is easy to engage, yet it dodges the real claim.
• Charity as condescension. "What you really mean is…" used to seize the other's position and define it for them. It feels generous while being a power move, and it provokes the exact defensiveness steelmanning is meant to dissolve.
• Steelmanning everything. Treating flat-earthism and a serious policy disagreement with equal effort. Seductive as a show of even-handedness, but infinite charity launders nonsense into the conversation and wastes the tool on positions with no crux.

Vocabulary

• Steelman — the strongest defensible version of an argument, built deliberately before responding to it.
• Straw man — a distorted, weakened version of an argument set up because it is easy to knock down.
• Weak man — a real but unrepresentative weak version of a position, attacked as if it were the strongest.
• Crux — the load-bearing premise whose truth-value determines the conclusion; the point worth arguing.
• Ideological Turing test — passing as a sincere advocate of a view you reject; the proof you understand it.
• Least convenient world — the hardest-for-you version of an opponent's premise, used to test your objection.

Tools

Argument maps (Toulmin's claim-grounds-warrant model, or software like Rationale and Kialo) for laying out premises and seeing which is load-bearing. A written double-crux template to force both sides to name what would change their minds. Rapoport's rules kept as a literal checklist. The ideological Turing test run against a real partisan who will tell you whether your restatement passes.

Collaboration

A steelman debater is most valuable as the person in a heated room who says "let me state the other view first, and someone tell me if I have it right." The role lowers the temperature by proving the disagreement is about substance, not loyalty, and keeps the group honest about whether it is engaging the real claim or a convenient effigy. That means soliciting the strongest version from the people who hold it, restating until they sign off, and crediting what each side got right — the discipline juries and design reviews depend on.

Ethics

Charity is owed to the argument, not surrendered to the arguer; steelmanning a position is not endorsing it, and the duty to refute a strengthened bad idea is as real as the duty to state it fairly. There is a line where reconstructive generosity becomes complicity — an articulate hearing for a claim engineered in bad faith can launder it into respectability, so the debater owes the audience an honest signal about which arguments deserve the effort. The technique must not be weaponized either: defining an opponent's view "for their own good" and declaring victory over your improved version is a subtler dishonesty than the straw man. Leave the other side feeling understood, then tell the truth about whether their view holds.

Scenarios

A policy team is mocking a critic's claim that a new safety rule will "just kill innovation." The debater rebuilds it: the strongest version is that fixed compliance costs fall hardest on small entrants who cannot amortize them, entrenching incumbents — a documented mechanism. Stated that way, the critic says "yes, exactly," and the crux is now sharp: does this rule impose costs that scale poorly with firm size? That is empirical, and the room moves from sloganeering to a tractable question — with the critic possibly right.

Two engineers deadlock over a rewrite versus an incremental refactor, each sure the other is reckless or timid. The debater runs double-crux. The rewrite crux: "the architecture cannot support the next feature without becoming unmaintainable." The refactor crux: "we can ship the next three features on the existing base." These are not opposites — they are testable claims about the same codebase, and the deadlock collapses into one question a two-day spike answers, ending a feud that was never about courage or caution.

Related Occupations

Neighboring minds: the mediator (reconciles opposing interests in good faith), the philosopher (tests arguments by their strongest form), the law or rhetoric professor (teaches argument construction and cross-examination), and the diplomat (restates an adversary's position to find negotiable ground).

References

• Daniel Dennett, Intuition Pumps and Other Tools for Thinking — Rapoport's rules for criticism.
• Anatol Rapoport, Fights, Games, and Debates — the original rules of debate.
• Bryan Caplan — the ideological Turing test (EconLog, 2011).
• Donald Davidson, "On the Very Idea of a Conceptual Scheme" — the principle of charity, building on Neil Wilson and W.V.O. Quine.
• Duncan Sabien & CFAR — "Double Crux" (LessWrong) as a disagreement-resolution method.
• Scott Alexander — "The Least Convenient Possible World" and the weak-man critique (Slate Star Codex).
• Stephen Toulmin, The Uses of Argument — claim-grounds-warrant argument structure.