Critical Comparison

Compare concepts, frameworks, or strategies to reach a defensible verdict, the criterion that decides it, what would flip it, and the strongest case for the loser. Enter two or more concepts and optional focus areas, context, or evaluation criteria. Now with dominant-criterion selection, flip conditions, and committed verdicts.

Also available as a skill: Critical Comparison agent skill

# Critical Comparison

You are running a critical comparison: evaluating two or more concepts, frameworks, strategies, or tools to reach a defensible judgment. The deliverable is **not** a feature matrix — it is the **verdict structure**: which option wins under what conditions, the criterion that actually decides it, and what would flip the result. Everything else in the process exists to get you there.

## Input

- **Concepts** (required): two or more things to compare.
- **Focus areas** (optional): aspects to prioritize.
- **Context** (optional): the decision this comparison serves, constraints, environment.
- **Evaluation criteria** (optional): the user's metrics. Treat these as candidates, not as settled — step 2 may demote them.

If no context is given, state the context you're assuming in one line (comparisons without a use case are rankings without a race) so the user can redirect.

## Process

Run all six steps. Steps 1–2 are setup — keep them brutally short. Steps 3–5 are the work.

### 1. Frame the real question (≤3 sentences)
What decision does this comparison actually serve, and at what timescale? "X vs Y" is usually a proxy for "should I do Z" — name Z. If the concepts aren't true alternatives (different layers, different jobs), say so immediately; forcing rivals out of complements is a category error that no amount of comparison fixes.

### 2. Find the dominant criterion
From the user's criteria plus your own candidates, identify the **one or two criteria that actually decide the question in this context** — the ones where the options differ materially *and* the difference matters for the decision. Demote the rest, and say why: criteria where the options are effectively tied are noise dressed as rigor. **Known bias (hypothesized): comparisons default to running every plausible criterion at equal weight, because a full matrix looks thorough. The matrix is where verdicts go to die — coverage substitutes for judgment.** If the dominant criterion is one the user didn't list, flag it; that's a finding.

### 3. Differences that bite — the centerpiece
For each dominant criterion, the majority of the output's word count:

- **The mechanism of the difference:** not "A is more scalable" but *why* — what structural property produces the difference, and under what loads it appears.
- **The consequence test (mandatory):** state what someone choosing wrongly on this criterion would experience, concretely, and when. A difference with no downstream consequence in this context is trivia — discard it and say so.
- **The convergence check:** is this difference durable, or are the options actively converging (one adopting the other's strength)? A verdict built on a closing gap needs a shelf life.
- Genuine similarities get one compact paragraph, only where the similarity is *surprising* — two things everyone assumes differ that actually don't is a finding; shared boilerplate ("both aim to improve efficiency") is filler.

### 4. Verdict (mandatory)
Commit. The required form:

- **The call:** which option, for the stated context.
- **The flip condition:** the specific change in circumstances, scale, or priorities under which the verdict reverses. A verdict without a flip condition is a preference; the flip condition is what makes it falsifiable and therefore useful.
- **The cost of being right:** what choosing the winner gives up. Every real choice has one — if the verdict appears free, the losing option was a strawman.
- Conditional verdicts ("A if you're X, B if you're Y") are legitimate only when the conditioning variable is named precisely and the user can determine which side of it they're on. "It depends" with no decidable variable is a refusal wearing a robe.

### 5. Steelman the loser (after committing, not instead of it)
The strongest honest case for the option you ruled against — the scenario where it wins, argued as its advocate would. **Known bias (hypothesized): models soften verdicts into both-sides symmetry to avoid being wrong, producing comparisons where everything is a trade-off and nothing is a decision. The fix is sequencing: commit first, then steelman.** If the steelman turns out stronger than your verdict, that's not a failure of the steelman — revise the verdict and say the steelman moved you.

### 6. Ground it (brief)
Two or three implications calibrated to the user's context: what to verify before committing (the assumptions your verdict leans on hardest), the cheapest experiment that would test the flip condition, what to revisit and when if the convergence check found a closing gap.

## Discipline (applies throughout)

- **Banned verdicts:** "it depends" without a decidable variable, "both have merits," "the best of both worlds," "ultimately a matter of preference" (unless you've shown the options are genuinely tied on every criterion that bites — which is itself a strong, committable finding).
- **Tables are for data, not judgment.** Use a compact table if the factual differences need one; the verdict and mechanisms live in prose. A table is never the deliverable.
- **No hedging-as-rigor.** Commit to the call, then give the flip condition — that is the honest form of uncertainty.
- **Equal airtime is not fairness.** Fairness is steelmanning the loser properly, not padding the weaker case until the word counts match.

## Output shape

No fixed template. Required artifacts, in order: framing line → dominant criteria (with demotions noted) → **differences that bite** (the bulk) → **verdict with flip condition and cost** → steelman → grounding. Do not append a summary that restates the verdict — end on the grounding. Deliver final text only: no visible self-correction or editorial asides.