zfc
Zero Framework Cognition Principles
When & Why to Use This Skill
The Zero Framework Cognition (ZFC) Principles skill provides a specialized architectural framework for building advanced AI agents. It focuses on 'pure orchestration,' where the application acts as a deterministic, safe shell that delegates all complex reasoning, planning, and decision-making to the LLM. This approach minimizes brittle, hard-coded logic and maximizes the flexibility and intelligence of AI-driven systems while maintaining strict safety guardrails and observability.
Use Cases
- Architecting Scalable AI Agents: Designing agentic systems that separate mechanical execution (IO, file handling) from cognitive tasks (planning, selection) to ensure cleaner codebases.
- Implementing Safety and Compliance Guardrails: Creating 'thin shells' around AI reasoning that enforce budget caps, rate limits, and schema validation without interfering with the AI's decision-making logic.
- Standardizing Agentic Workflows: Applying the ZFC-compliant pattern (Gather Context -> Call AI -> Validate -> Execute) to build consistent and predictable AI-powered automation tools.
- Refactoring Brittle Heuristic Systems: Replacing complex, rule-based decision trees and hard-coded 'if-else' logic with LLM-driven reasoning for better handling of edge cases and semantic complexity.
| name | zfc |
|---|---|
| description | Zero Framework Cognition Principles |
ZFC (Zero Framework Cognition) Principles
Core Architecture Principle: This application is pure orchestration that delegates ALL reasoning to external AI. We build a “thin, safe, deterministic shell” around AI reasoning with strong guardrails and observability.
✅ ZFC-Compliant (Allowed)
Pure Orchestration
IO and Plumbing • Read/write files, list directories, parse JSON, serialize/deserialize • Persist to stores, watch events, index documents
Structural Safety Checks • Schema validation, required fields verification • Path traversal prevention, timeout enforcement, cancellation handling
Policy Enforcement • Budget caps, rate limits, confidence thresholds • “Don’t run without approval” gates
Mechanical Transforms • Parameter substitution (e.g., ${param} replacement) • Compilation • Formatting and rendering AI-provided data
State Management • Lifecycle tracking, progress monitoring • Mission journaling, escalation policy execution
Typed Error Handling • Use SDK-provided error classes (instanceof checks) • Avoid message parsing
❌ ZFC-Violations (Forbidden)
Local Intelligence/Reasoning
Ranking/Scoring/Selection • Any algorithm that chooses among alternatives based on heuristics or weights
Plan/Composition/Scheduling • Decisions about dependencies, ordering, parallelization, retry policies
Semantic Analysis • Inferring complexity, scope, file dependencies • Determining “what should be done next”
Heuristic Classification • Keyword-based routing • Fallback decision trees • Domain-specific rules
Quality Judgment • Opinionated validation beyond structural safety • Recommendations like “test-first recommended”
🔄 ZFC-Compliant Pattern
The Correct Flow
Gather Raw Context (IO only) • User intent, project files, constraints, mission state
Call AI for Decisions • Classification, selection, composition • Ordering, validation, next steps
Validate Structure • Schema conformance • Safety checks • Policy enforcement
Execute Mechanically • Run AI’s decisions without modification