Critic Agent Configuration

A generic QA prompt — "review this for quality" — catches surface-level issues at best. Spelling mistakes, obvious formatting problems, maybe a factual error if it is egregious. But the subtle problems slip through. The argument that sounds logical but rests on a flawed assumption. The chart that is technically accurate but visually misleading. The recommendation that contradicts a constraint mentioned four pages earlier.

Specialized critic agents solve this. Instead of one generic reviewer, you configure critics with domain-specific review criteria. A financial critic checks for calculation errors, consistency between tables and narrative, and appropriate use of caveats. A legal critic checks for ambiguous language, missing disclaimers, and regulatory compliance. A brand critic checks for tone consistency, visual identity adherence, and messaging alignment.

The configuration is in the prompt. You are not writing code. You are writing a detailed job description for a very focused reviewer. Here is what a financial critic prompt looks like: "You are a financial review specialist. Evaluate this document on the following dimensions: (1) Are all calculations correct? Verify every number by recomputing from source data. (2) Do narrative claims match the data in tables and charts? Flag any contradiction. (3) Are projections labeled as projections and not presented as facts? (4) Are appropriate caveats included for forward-looking statements? (5) Is the level of precision appropriate — no false precision on estimates, no rounding errors on actuals."

That prompt creates a critic agent that catches financial-specific errors a generic reviewer would miss. The more specific the review criteria, the more specific the catches. Generic criteria produce generic reviews. Domain criteria produce domain-grade quality assurance.

1. 1. Identify the Domain What kind of deliverable is being reviewed? Financial analysis, legal document, marketing copy, technical specification? The domain determines the review criteria.
2. 2. Write Domain-Specific Criteria List 4-6 specific evaluation dimensions. Not "is it good?" but "does every claim have a cited source?" and "do all percentage calculations match the underlying numbers?" Each dimension should be verifiable.
3. 3. Define the Scoring System Give the critic a numeric scale per dimension and a threshold for approval. "Score each dimension 1-10. Approve if all dimensions are 7 or above. If any dimension is below 5, flag as critical and return revision instructions."
4. 4. Save as a Skill Once your critic prompt is refined through 2-3 uses, save it as a skill. "Financial Document QA" becomes a reusable tool. The next time you produce a financial deliverable, the critic is already configured.