OWNED SKILL · STEELMAN

Your AI agrees
with you. That's
worthless.

Once you've decided you want something, you're the worst-positioned person alive to build the case against it. So the AI does what you can't: it argues both sides at full strength, with zero ego in the outcome.

The AI argues. You judge.
No ego in the outcome — that's the point.

say "steelman this"
CASE FOR

Argued as its most capable proponent would.

The real upside. The cost of not doing it.

No faint praise.

CASE AGAINST

Equal strength. No strawmen.

The best objection. The failure mode.

The thing you're motivated not to see.

↓  THE CRUXES
The few load-bearing assumptions that, if flipped, flip the answer.
THE AI IS THE ADVOCATE. YOU ARE THE JUDGE.
↓  WHY IT EXISTS

The problem

Ask an AI whether your plan is good and it will tell you your plan is good. That agreement is cheapest exactly when the stakes are highest — you've already decided, you're invested in the conclusion, and now the one thing in the room that could talk you out of it is nodding along. That's not a flaw in you; it's how minds work. It's just useless.

The turn

Steelman flips the role. The AI is the advocate; you are the judge. It builds the case for the way its most capable proponent would, then the case against at equal strength — the best objection, the failure mode, the thing you're motivated not to see. No strawmen. Then it names the cruxes: the load-bearing assumptions that, if flipped, flip the answer.

It resolves every fact it can itself — codebase, web, memory — and only taps you where your private knowledge, risk tolerance, or values are the deciding factor. One crux at a time, not a barrage. Then it stops and hands you the verdict. It may offer a labeled tentative lean; it never substitutes that for your call.

When the opposition feels soft: one mind arguing with itself can produce a polite, symmetric both-sides where nothing lands. On calls that matter, it escalates — spawning an independent subagent whose only mandate is to kill the idea. A prosecutor with no obligation to be fair. Its independence is the whole value.

When to fire it

Hard-to-reverse or high-stakes calls:

Works in both directions: before a decision it's a pre-mortem (catch it before you pay for it); after something's built it's a red-team. Skip trivial and reversible calls — proportionality. For small ones, the in-context pass is the right amount of effort.

How it's wired in

A steelman won't self-invoke — which is exactly why it's a standing rule rather than a skill you're trusted to remember. AGENTS.md §3 (Planning & Verification) makes it mandatory before any hard-to-reverse or high-stakes decision.

# AGENTS.md §3 — Planning & Verification
- Steelman consequential calls: before committing to a hard-to-reverse
  or high-stakes decision ... invoke the `steelman` skill — it argues the
  strongest case *for* and *against*, surfaces the cruxes, then I judge.
  A steelman won't self-invoke — that's why it lives here as a standing rule.

r2-sdlc Phase 2 (Design) hard-codes the same beat: if the approach rests on a hard-to-reverse choice, steelman it before the design ever reaches the user's redline — so what gets reviewed is already pressure-tested. Skipped for small, mechanical changes.

Part of Rascal's AI setup · one atom, justifying its own existence.