Model tier guide — which models to pick for each agent type
Each model in the catalog is grouped into a tier reflecting its intended use. Pick the tier that matches the agent's job — fast models for scoring, balanced for analysis, premium for writing. Assignments here override the global default per agent.
Default Z.AI subscription model. Used by legacy refresh paths and as the catch-all when no agent override exists.
Low-latency, low-cost models. Recommended for evaluators, validators, and small judgment tasks (image review, alt-text, placement) where speed and price matter more than narrative quality.
Mid-tier reasoning models. Recommended for the analysis and planning phases — content_analyzer, keyword_researcher, blueprint, link_planner — where reasoning over context helps but a frontier model is overkill.
Top-quality writing models. Reserve for the writing phase (content_writer, polisher, composer) where output quality is king.
OpenRouter free router. Useful for fallback chains and testing — quality and availability vary.
Backup providers kept available for diversification or A/B comparisons. Not recommended as primary assignments.
Evaluator Agents
Score content quality across multiple axes. Run in parallel via fan-out (Bus::batch). Model, system prompt overrides, and custom instructions are saved as immutable revisions with history.
On-page SEO signals: keywords in headings, density, meta/alt text, internal linking
Revision history (1)
3 days ago · Initial seed from hardcoded configuration
Content quality: specificity, originality, E-E-A-T signals, topic coverage
Revision history (1)
3 days ago · Initial seed from hardcoded configuration
HTML structure: heading hierarchy, intro/conclusion, formatting, HTML hygiene
Revision history (1)
3 days ago · Initial seed from hardcoded configuration
Keyword strategy: gap coverage, intent alignment, cannibalization risk, distribution
Revision history (1)
3 days ago · Initial seed from hardcoded configuration
Native fluency and script purity: correct script usage, consistent register, and natural idiomatic expression free of translation artefacts
Revision history (1)
3 days ago · Initial seed from hardcoded configuration
Reading ease: sentence complexity, paragraph length, vocabulary level, Flesch-Kincaid approximation
Revision history (1)
3 days ago · Initial seed from hardcoded configuration
Link quality: anchor diversity, internal/external ratio, nofollow usage, broken link indicators
Revision history (1)
3 days ago · Initial seed from hardcoded configuration
Google E-E-A-T signals: Experience, Expertise, Authoritativeness, Trustworthiness
Revision history (1)
3 days ago · Initial seed from hardcoded configuration