When the LLM becomes the Designer: Vellum and the Maturity of AI-Driven Creative Workflows
It's not just a wrapper. It's not a tool that sends prompts to a model and waits for a response. Vellum is a complete creative workflow orchestration platform built around new-generation models that demonstrate what "Astra-class" systems can do when they step out of the realm of chatting.
The result is surprising: an environment similar to Figma with vector editing, layer management, annotation tools, and real-time collaboration, all orchestrated by a single LLM context. The difference from classic AI assistants isn't in computational power — it's in how the model manages the state of the project.
The Architecture of Recursive Memory
Vellum's secret isn't the model. It's the mechanism of contextual anchoring.
While most AI tools apply a static prompt to each iteration, Vellum maintains a recursive memory structure that links every modification to the project's global context. When the model generates a new modification to a layer or component, the system doesn't "lose the thread." It uses a system of semantic pointers to anchor every action to a previous state, creating a chain of dependencies that the model can navigate without context hallucinations.
This is the difference between automation and co-creation. The first executes instructions. The second understands the structure of the project and anticipates the consequences of its modifications.
Why This Matters for SMEs
For those who develop software or manage creative workflows, Vellum demonstrates three concrete things:
1. New-generation models are mature enough for non-deterministic tasks. Figma isn't a repetitive task. It requires aesthetic judgment, understanding of visual hierarchies, anticipating how a designer will interact with the product. If a model can manage this, it can manage any creative workflow.
2. Zero training cost means immediate accessibility. SMEs don't need to build proprietary datasets to automate prototyping. They can use open-weight models or standard APIs, integrating them into existing pipelines.
3. Human-AI collaboration is now structured, not accidental. Vellum doesn't replace the designer. It provides a layer of scaffolding that reduces ideation time from days to hours. The model generates options; the human approves, modifies, directs.
Practical Insight: Implement Contextual Anchoring in Your AI Tool
If you're building an AI app today, avoid the "prompt → response" pattern. Instead, implement a