Steelman

The fastest way to be wrong with confidence is to argue against the weakest version of the other side. This skill does the opposite: it builds the strongest honest version of the opposing view first — the one its smartest advocate would actually defend — and only then engages it. Along the way it does something most "see both sides" advice skips: it figures out what kind of disagreement you're in, because factual disagreements and values disagreements need completely different handling, and treating one as the other wastes everyone's time.

Two jobs

1. Steelman the position you're inclined to reject — state it at full strength, from inside its own assumptions, with the best available support.
2. Classify the disagreement — factual (data could resolve it) vs. values/priors (data can't) — so you know whether to gather evidence, or to stop pretending evidence will settle it.

Job 1: Build the steelman

A real steelman passes this test: its advocate would say "yes, that's my view — better than I put it." If they'd say "that's not what I mean," you built a strawman with better manners.

Rules:

• Argue from inside its assumptions, not yours. Grant its premises for the length of the argument and see how far they honestly reach.
• Use the best available evidence for it, not the weakest. Find the strongest study, the sharpest mechanism, the most capable proponent — and represent that.
• Fix its avoidable mistakes. If the position has a sloppy version and a rigorous version, steelman the rigorous one. Defeating the sloppy version proves nothing.
• State what would have to be true for the position to be correct. This is the hinge — it tells you exactly what to test.
• No poisoning. Don't smuggle in "of course, this ignores…" while building it. Build it clean; critique comes after, clearly separated.

Job 2: Classify the disagreement

Once both sides are stated at full strength, name what kind of gap separates them:

• Factual disagreement — the sides make different claims about how the world is, and some observation could in principle move them. ("Will the trend continue?" "Does the drug reduce mortality?") → There's a resolution path. Name the specific evidence that would settle it, even if you don't have it yet.
• Values / priors disagreement — the sides read the same facts and still diverge, because they weight outcomes differently or hold different deep priors about how the system works. (One expert's "misalignment at scale is near-inevitable" vs. another's "safety can be engineered incrementally" — same literature, different prior about the structure of the problem.) → Data won't settle this. Report the disagreement as real and possibly permanent; don't pretend one more study will resolve it. For the values case, route to the value-frames skill.

The tell for a values disagreement: both sides are informed, both have looked at the same evidence, and the gap hasn't narrowed as evidence has accumulated. Narrowing gaps are factual; stubborn gaps that track priors are not.

Don't average the disagreement away

When credentialed people disagree by a wide margin, the spread is the finding — don't collapse it to a midpoint that no one holds and that manufactures false consensus. Report the distribution and the outliers, name where the median and mean sit, and say plainly whether the disagreement is narrowing (factual, wait for data) or stable (priors, won't resolve). A single "the answer is 7%" is almost always less honest than "estimates range from ~1% to ~99%, cluster around 5–20%, and haven't converged in three years — which itself suggests the split is about priors, not evidence."

Procedure

1. Identify the position you're inclined to reject (or the side of a decision you're not on).
2. Build its steelman to the "advocate would endorse it" bar.
3. State what would have to be true for it to be right.
4. Classify the disagreement: factual (→ name the resolving evidence) or values/priors (→ name the value or prior at stake; consider value-frames).
5. Now give your critique or your call — clearly separated from the steelman, and honest about which parts of the steelman survive.
6. If experts disagree widely, report the distribution, not an average.

Output template

POSITION STEELMANNED: <the view, stated so its advocate would endorse it>
  Best support: <strongest evidence/mechanism for it>
  Would have to be true: <the hinge conditions>

DISAGREEMENT TYPE: factual | values/priors
  If factual → resolves if: <the specific observation that would settle it>
  If values → the value/prior at stake: <what the sides actually differ on>

MY ASSESSMENT: <critique or decision — separated, and saying which parts
  of the steelman still stand after scrutiny>

Example

Input: "Should we ban the open-weight release of frontier models? I think obviously yes."

Steelmanned (the no-ban side you're inclined to reject):

POSITION STEELMANNED: Open weights should stay legal. Open release is the main
  force pushing capability out of a five-firm oligopoly, giving researchers,
  the Global South, and safety auditors access they'd otherwise be denied.
  Closed models aren't safer — they're unauditable; the strongest interpretability
  and red-team work depends on open access. Bans mostly disarm the people
  studying the risk, not the state actors who'll obtain weights anyway.
  Best support: open-weight models already carry a large share of real developer
  usage; independent safety research overwhelmingly runs on open models.
  Would have to be true: marginal misuse uplift from open weights is smaller
  than the safety/access benefit — an empirical claim, currently contested.

DISAGREEMENT TYPE: mixed. The misuse-uplift magnitude is FACTUAL (resolves if a
  documented real-world open-weight-enabled incident, or its sustained absence,
  moves the number). The "distributed access vs. centralized control" preference
  underneath is VALUES — even with the facts fixed, people weigh diffusion-of-
  power against concentration-of-risk differently.

MY ASSESSMENT: "Obviously yes" was overconfident. The factual half is genuinely
  unresolved, and the values half won't be settled by a study at all.

The point isn't to flip your view. It's that "obviously" almost never survives a real steelman — and now you know exactly which part is an evidence question and which is a values question.

Gotchas

• A steelman is not agreement. Building the best version of a view you still reject is the whole exercise. You can steelman and then dismantle it.
• Don't steelman bad faith into good faith. Represent the strongest honest version — not a laundered version of a position whose real appeal is dishonest.
• If you can't build a steelman, you don't understand the disagreement yet. That's information: go learn the other side before critiquing it.
• Pairs with value-frames (for the values half) and grade-the-evidence (for the factual half).