CRIT — the Critical Inquisitive Template — is a prompt template for making a language model evaluate an argument instead of just reacting to it. It comes out of Edward Y. Chang's work on prompting LLMs with the Socratic method, and it shows up downstream in multi-model debate frameworks, where something has to decide which participant actually argued well.
People find it while looking for a principled way to score AI output. It is a good thing to know about. It is also frequently pointed at the wrong artifact, which is the part worth being clear on.
What CRIT actually does
Give CRIT a document and it produces a validation score. The mechanism is decomposition:
- Extract the conclusion. What is this document actually claiming?
- Extract the reasons. What supporting claims are offered for that conclusion?
- Score each reason-to-conclusion link twice. Once for argument validity — does this reason, if true, actually support the conclusion? Once for source credibility — is there reason to believe the supporting claim itself?
- Combine. The per-argument scores roll up into a single validation score for the document.
The two-factor split is the clever part. A perfectly valid inference from a made-up premise and a shaky inference from an impeccable source are both weak, and they are weak in different ways that a single "quality" number would smear together. Scoring validity and credibility separately keeps them distinguishable.
The Socratic framing is not decoration either. Chang identifies ten strategies — definition, generalization, induction, elenchus, hypothesis elimination, maieutics, dialectic, recollection, irony, and analogy — and CRIT draws on specific ones for specific jobs: definition, elenchus, and dialectic for critical reading; maieutics and counterfactual reasoning for critical thinking. Elenchus is the load-bearing one. It is the move where you take a claim seriously enough to try to derive a contradiction from it.
Why the decomposition is the real lesson
Strip away the Greek and CRIT is making one argument: do not ask a model for a holistic verdict.
"Rate this essay 1 to 10" gets you a number that correlates with length, confidence, and vocabulary. "Extract the claim, list the reasons, score each link" gets you a number you can interrogate — and, more usefully, a trail showing which specific reason dragged the score down.
That is the same reason a rubric beats a vibe. Both replace one unanswerable question with several answerable ones. We made that case for instruction quality in stop grading prompts on vibes; CRIT makes it for argument quality. Different artifact, identical insight.
Where CRIT stops
Here is the mistake we see: someone reads about CRIT, likes it, and points it at their system prompt.
It will return a number. The number will not mean much, because a prompt is not an argument. CRIT's whole machinery assumes a document with a conclusion and supporting reasons. An instruction has neither. It has directives. There is nothing to validate, no premise whose credibility is in question, no inference to check.
And the ways instructions actually fail are not the ways arguments fail:
- Ambiguity. Two readings of the same sentence, both defensible, that produce different behavior.
- Unbounded behavior. No statement of what the model should refuse or leave alone.
- Missing output specification. No format given, so the format drifts.
- Internal contradiction. Line 4 and line 40 asking for different things.
CRIT is not built to see any of those. That is not a deficiency in CRIT — it is aimed at a different target. Scoring an instruction is what the ten-dimension rubric is for. Scoring a generated argument is where CRIT-style decomposition earns its keep.
The caveat that applies to both
Whatever template you use, a model scoring against it is not a ruler. Run the same evaluation twice on unchanged text and the score moves — in our measurements, typically around a point, but with wildly different spread depending on the artifact. We have seen a standard deviation of 3.06 on one and 0.82 on another: same rubric, same model, nearly four times the noise.
This applies to CRIT exactly as much as to a rubric score. A validation score of 78 against 76 on two drafts is not evidence that the second draft argues worse. Before you believe any delta:
- Score the baseline more than once. One run is a sample, not a value.
- Repeat the candidate too. Otherwise you are comparing two lucky draws.
- Ask whether the gap exceeds the noise. That is a hypothesis test, and it has an answer — see statistical significance for prompts.
Decomposition makes a score more interpretable. It does not make it deterministic. Those two get conflated constantly, and the result is teams shipping "improvements" that were measurement noise wearing a Greek name.
How they compose
The useful framing is that the two grade different halves of the same system:
- A rubric grades the instruction — before it runs, when it is cheapest to fix.
- CRIT-style decomposition grades the output — the reasoning chain, the research summary, the argument the model actually produced.
If your agent writes analyses that have to hold up, you want both: a graded instruction going in, and structured critique of what comes out. Grading only the output means relitigating the same weakness every run, because the instruction that caused it never changed. That is the case for auditing the prompt before you run the agent.
Learn CRIT for the decomposition. Apply it to arguments. Do not ask it to tell you whether your prompt is any good.