15 Foundational Concepts Behind Codex Working in Figma

When Codex reads Figma, finds components, changes layouts, and creates pages, it is running a continuous information-processing and tool-execution chain.

These fifteen concepts explain what happens between a designer's request and a real change in the Figma file.

Token

The basic unit used to encode text, node names, tool descriptions, code, results, and errors before they enter the model context.

Context

Everything the model can currently reference: the request, conversation, system rules, Skills, tool definitions, Figma nodes, screenshots, and previous results.

Embedding

A numerical representation used to compare semantic similarity, helping the Agent retrieve components even when the user's wording differs from asset names.

Transformer

The model architecture that relates requirements, design nodes, tools, Skills, and earlier operations across a long context.

Attention

The mechanism that assigns different importance to information in context, such as component instances for reuse or Auto Layout properties for a two-column request.

Autoregressive generation

The model predicts the next token or next action step by step, which is why Codex appears to think and act incrementally.

Tool Calling

The structured instruction through which model decisions become real reads and writes performed by external tools.

Tool Schema

The contract that specifies parameter names, types, required fields, and allowed values so natural-language intent can become an executable operation.

Agent Loop

The repeated cycle of observing the environment, selecting an action, calling a tool, reading the result, and deciding what to do next.

Multimodal visual encoding

The process that turns screenshots into representations of regions, text, color, spatial relationships, hierarchy, clipping, and visual imbalance.

Grounding

The process of mapping phrases such as 'the second metric card on the right' to one real node ID using names, hierarchy, position, component data, and screenshots.

Context-conditioned control

Skills place rules into the current context, changing the likelihood that Codex searches existing components, uses variables, preserves instances, and verifies screenshots.

Sampling and uncertainty

Several actions may appear plausible at each step. Context, tool order, retrieved nodes, sampling, and earlier choices can make identical requests follow different paths.

Feedback loop

After an operation, Codex reads structural and visual feedback, locates the remaining difference, and continues revising rather than assuming success.

Idempotency

A reliability property ensuring repeated execution does not create duplicate pages, components, or variables and instead updates or skips existing objects.

One sentence for the whole mechanism

Codex combines the request, tool definitions, Figma nodes, Skills, and earlier results in context; predicts the next action; calls an external tool to change Figma; then uses the returned nodes and screenshots to continue the observation, action, and feedback loop.