Viktor.comAn AI coworker that actually does the work
Promoted
memi
The AI agent harness for product design teams
124 followers
The AI agent harness for product design teams
124 followers
A macOS workbench where Claude, Codex, and Hermes run on your specs, research, and Figma files.
memi
Routing multiple agents against the same Figma source is the right call. Text specs drift from design intent the moment a file changes. It's a translation layer that doesn't survive iteration. When Claude and Codex generate diverging implementations from the same component spec, how do you surface and resolve those conflicts?
memi
@anand_thakkar1 Great question. The way we think about it is: Figma stays the source of truth, and agent outputs become proposals against that source, not separate truths. memi keeps the shared context around component specs, tokens, source frames, receipts, and prior accepted changes. When Claude and Codex diverge, the goal is to surface the delta clearly: what changed, which design intent it maps to, and where it conflicts. The designer can accept, reject, or revise, and that resolution becomes part of the project memory for the next run. So conflict resolution is less “average two agents” and more “keep design intent anchored while agents produce inspectable diffs.”
Love the vision of an AI that understands your patterns and taste. I wonder how does memoire actually learn and quantify something as subjective as a designer's aesthetic preferences over time?
memi
@crystalmei I love this question because taste is the whole point. We do not try to turn a designer’s aesthetic into one magic score. memi builds a living taste profile from repeated signals: accepted vs rejected generations, recurring token choices, component patterns, spacing/type/color preferences, reference boards, Figma usage, and project-specific decisions. Over time it can say things like “this team tends to prefer denser layouts, sharper radii, quieter contrast, and editorial type” and use that as context for critique or generation. The quantification is really a memory layer plus receipts, so subjective taste becomes something the agent can reference and improve against without pretending it is objective.
Is Figma required? Where I struggle with design systems in AI-native development is maintaining and verifying against a single source of truth.
memi
@jgilbertson47 nope not required at all!
Love this idea. Taste is one of the hardest things for AI to learn