Most AI workflows send every task to one model and hope for the best. The CYW architecture treats different models as specialists — and keeps a human in charge of deciding who handles what.
Perplexity is fast and grounded in sources. Claude reasons through complexity and writes with nuance. Gemini catches jargon and improves accessibility. DeepSeek structures logic and generates clean code. ChatGPT synthesizes and formats for publication.
Routing is the skill of knowing which model to use — and when. Without a routing layer, you're either picking one model for everything or making ad-hoc decisions that are hard to repeat or explain.
CYW assigns each AI model a role in a five-stage pipeline. A task can enter at any stage — and a human decides whether to move it forward, send it back, or stop.
VLAN is borrowed from network architecture — Virtual Local Area Network. In networking, VLANs separate traffic by role. In CYW, they separate AI work by function. Each model operates in its lane; the human controls the handoffs.
Entry point for tasks that need grounding in real, citable information. Use when the task requires current data, citations, or verifiable facts before any generation begins.
The generative layer. Handles complex reasoning, long-form writing, nuanced framing, and tasks that require holding a lot of context. Strong on voice and narrative architecture.
The audit layer. Routes here when the goal is review — strip jargon, improve readability, check that a piece is accessible to its intended audience. Not generation; refinement.
The structure layer. Use for code generation, logical restructuring, argument integrity checks, and tasks where correctness and consistency matter more than style.
The output layer. Integrates work from earlier stages into a coherent, publication-ready artifact. Strong at formatting, cross-referencing, and assembling multi-source outputs.
The Prompt Router scores each task across six dimensions. Scores are multiplied by each model's affinity for that dimension — the model with the highest total affinity is the suggested starting point.
| Dimension | What it measures | Best-fit model |
|---|---|---|
| Complexity | Single-fact lookup → multi-step reasoning | Claude |
| Domain specificity | General → highly specialised field | Claude / DeepSeek |
| Output structure | Freeform prose → structured / executable | DeepSeek |
| Latency sensitivity | Async / background → real-time | DeepSeek |
| Verification need | Low-stakes → requires sourcing / citations | Perplexity |
| Review / Audit | Net-new creation → clarity or jargon audit | Gemini |
Three browser-based tools that put the CYW architecture into practice. No accounts. No API keys. Open and use.
15-rule detection framework for AI-generated responses. Flags fabricated citations, precision abuse, logical inconsistency, and instruction drift.
Open tool →Tone-reactive AI counterpart across five scenarios. Push too hard and they push back; be collaborative and they open up.
Open tool →Paste a task. The router scores it across six dimensions and suggests a starting model based on the VLAN architecture. Rationale first, recommendation second.
Open tool →Exploring Contruil for your organization? See the platform →