The "paste 200 words of context into every prompt" problem is exactly what kills agent workflows at scale. Marque moves the design identity to infrastructure, which is the right fix for that specific layer.

There's a parallel problem at the prompt content layer: even when you inject the context, if the prompt itself is flat prose the agent still interprets it inconsistently run to run. Structuring the injected context into named blocks (constraints, role, output format) makes the agent much more deterministic.

Built github.com/Nyrok/flompt for this, a visual prompt builder that decomposes instructions into semantic blocks before they reach the agent. Pairs well with something like Marque sitting at the infra level.