Steelman Before You Kill

When to use

• The user (or your own analysis) is about to reject an idea, a business concept, a strategy hypothesis, an architecture choice, or a vendor/tool.
• You are asked to "evaluate", "critique", "stress-test", "poke holes in", or "score" anything.
• You notice yourself pattern-matching an idea to a known failure class ("that's just X, and X fails").
• A prior evaluation ended in "no" and someone asks to revisit it.

When NOT to use

• Pure factual disputes with a checkable answer (a number is right or wrong — go check it; no steelman needed).
• Safety, legality, or hard-constraint violations. An idea that requires breaking a terms-of-service or a law gets no steelman; it gets a boundary statement.
• Trivial reversible decisions (naming a variable, picking a linter). Steelmanning everything is procrastination wearing a methodology costume.

Why this is a procedure and not a virtue

Pattern-matching to failure classes is usually right and occasionally catastrophically wrong, and you cannot tell which from the inside. Real case: a "niche storefront for X" idea was pattern-matched to "marketplaces are power laws, median revenue ≈ 0" — a rule that had been empirically confirmed across dozens of prior evaluations. The critique was correct and the idea died. But the steelman pass, run first, surfaced three billion-dollar companies whose founding shape was exactly the pattern being dismissed, plus a probe design costing under one week of effort that would have detected the difference between the median case and the outlier case. The main idea still died — but the probe design and one precedent's distribution insight were reused in the next three evaluations. The steelman is what converts "no" into "no, but THIS survives." Without it, a rejection destroys 100% of the information in the idea. With it, a rejection destroys only the dead part.

The procedure

Run steps 1–4 before writing a single sentence of critique. Ordering is the entire method: critique-first contaminates everything after it (see failure mode 2).

1. State the idea in its strongest form, not its stated form

The version presented to you is rarely the best version of itself. Rewrite it:

• Fix free parameters at their most plausible values, not their stated values. If the proposal says "weekly digest email product" and daily cadence is obviously stronger for the use case, steelman the daily version and say so.
• Name the concrete winning variant: the specific configuration, market segment, or parameter choice under which this idea wins. "A monitoring product" is not steelmannable; "a change-detection monitor for procurement notices, sold to the 8 firms that currently pay a human to check daily" is.
• If you cannot construct any variant that could win, say that explicitly — that itself is a finding, and it must survive step 4's audit like any other claim.

2. Find winning precedents — with numbers, or they don't count

Search for real instances where this shape of idea worked. The bar:

• Named entities with quantities: revenue, users, funding, growth rate, time-to-traction. "Some companies have done this" is not a precedent; "Company A reached $Nm ARR in 3 years starting from exactly this wedge" is.
• At least one precedent must match the idea's constraints, not just its shape. A precedent that needed $2M of capital does not steelman a plan with a $500 budget; find the bootstrapped instance or mark the precedent as shape-only.
• For each precedent, extract the discriminating condition: what was true for the winner that is false for the median failure? This is the single most valuable output of the whole exercise — it becomes the probe target in step 3.
• Two to four precedents. One is an anecdote; ten is padding.

3. Design the cheapest probe that would prove it

Specify the smallest, cheapest experiment that would move the idea from "plausible" to "validated" (or kill it cleanly):

• State cost (hours and money), duration, and the pre-registered pass/fail threshold — written down before the probe runs, so the result can't be reinterpreted afterward.
• The probe must test the discriminating condition from step 2, not general enthusiasm. If winners won because demand existed before supply, the probe measures demand independently of supply — it does not build the product and wait.
• If the cheapest honest probe costs more than the idea's plausible payoff, that IS the critique, and it's a much stronger one than any pattern-match.

4. Only now: audit

Write the critique. Rules of engagement:

• The critique must meet the same evidentiary bar as the steelman: cited failure precedents with numbers, base rates with sources, checked constraints. A cited steelman demolished by a vibes-based critique is a rigged trial in the other direction.
• The critique must engage the strongest variant from step 1, not the stated version. Killing the weak version of an idea proves nothing.
• Every steelman claim gets a verdict: survives, refuted (with the specific evidence), or untested — probe required.

5. Write the salvage line

End every negative verdict with an explicit inventory of what survives: precedents worth remembering, the probe design (often reusable for sibling ideas), the discriminating condition, any sub-component that works standalone. If the verdict is "no" and the salvage inventory is empty, you almost certainly skipped step 2 or 3 — go back.

Common failure modes

1. The token steelman. One vague complimentary paragraph ("there's real demand for this space") followed by five pages of critique. Wrong: the steelman has no named precedents, no numbers, no probe — it exists to be knocked down, which makes it a strawman with better posture. Right: the steelman section is comparable in length and citation density to the critique. If your critique cites 6 sources and your steelman cites 0, the trial was rigged before it started.

2. Critique-first anchoring. You spot the fatal flaw in the first 30 seconds, then "do the steelman" afterward for form's sake. Wrong: everything you generate after forming the verdict is unconsciously curated to support it — you will search for weak precedents and find them. Right: mechanical ordering. Steelman artifacts (precedent list, variant, probe) are fully written before the first critical sentence. If you already know the flaw, park it in one line at the top ("suspected flaw: X — to be tested in step 4") and proceed.

3. Pattern-match as verdict. "This is a [known failure class], therefore no." Wrong: the pattern gives you a base rate, not an outcome — and the entire economic value of evaluation lives in the off-base-rate cases. In the storefront case above, the base rate said "median ≈ 0" and was right; but three billion-dollar outliers shared a checkable discriminating condition the base rate was blind to. Right: pattern-match sets the prior; the steelman's precedents and probe determine whether this instance escapes it. Write the sentence "the base rate is X; this instance escapes it only if Y; the probe tests Y."

4. Steelmanning the median instead of the best. Evaluating the idea as the proposer stated it, mediocre parameters and all. Wrong: "as stated, this fails" is trivially true of most first drafts and teaches nothing. Right: step 1 exists precisely so the critique has to beat the idea at its best. "Even the strongest variant fails because Z" is a durable finding; "the draft fails" is not.

5. No salvage extraction. Verdict delivered, file closed, everything discarded. Wrong: across a long research program, the salvage lines compound into more value than the verdicts — probe templates get reused, precedent libraries prevent re-litigating settled questions, discriminating conditions become screening filters for future ideas. Right: step 5 is mandatory even (especially) on hard rejections.

Checklist

• Strongest variant written, with free parameters set to best values (not stated values)
• 2–4 winning precedents, each with named entity + numbers + discriminating condition
• At least one precedent matches the idea's actual constraints (capital, time, skills)
• Cheapest probe specified: cost, duration, pre-registered pass/fail threshold
• All steelman artifacts completed BEFORE the first critical sentence
• Critique cites evidence at the same density as the steelman
• Critique targets the strongest variant, not the stated draft
• Every steelman claim tagged: survives / refuted / untested
• Salvage line written: what survives the verdict, explicitly listed

Output template

## Steelman: <idea>
**Strongest variant:** <one concrete configuration that could win>
**Precedents:**
1. <Entity> — <numbers>. Discriminating condition: <what made it win>
2. <Entity> — <numbers>. Discriminating condition: <...>
**Cheapest probe:** <design> | cost: <hours/CHF> | pass if: <pre-registered threshold>

## Audit
<critique, same evidentiary bar, targeting the strongest variant>
**Claim verdicts:** <claim → survives / refuted(evidence) / untested>

## Verdict: <GO / NO / PROBE-FIRST>
**Survives regardless:** <precedent insights, probe design, reusable components>