Large Figma Feature

Extract a large feature breadth-first, then deepen only where evidence is needed. Preserve node IDs in every artifact so findings remain traceable.

Workflow

1. Call list_files, then get_metadata. Require an explicit fileKey whenever multiple files are connected.
2. Call get_selection. If nothing relevant is selected, use metadata and a shallow design context to identify the target section; do not fetch the full document tree.
3. Call get_design_context with depth: 1. Treat this as the feature inventory. Record top-level node ID, name, type, bounds, and child count.
4. Classify top-level nodes into entry, configuration, transient state, live state, result, modal, and supporting screen groups. Flag duplicate names and ambiguous variants.
5. Fetch individual top-level nodes with get_node, in small batches. Compact each response before showing it: retain node identity, bounds, meaningful component/layer names, unique visible text, state cues, and interaction clues. Discard repetitive vector geometry and duplicated styles.
6. Use get_screenshot only for visually ambiguous nodes. Prefer targeted screenshots over exporting the entire section. Use save_screenshots only when local image artifacts are an explicit deliverable.
7. Inspect the repository after the design inventory. Map screens and actions to routes, components, types, APIs, persistence, realtime events, and tests.
8. Reconcile Figma, user-provided notes, existing behavior, and code. Never silently choose between conflicting sources; report the conflict and its implementation consequence.
9. Produce only the requested artifact. Use the schemas in references/output-schemas.md for comprehensive analysis.

Context limits

• Never call get_document for a large page or get_node for the whole feature section when a shallow context can enumerate its children.
• Keep each detailed batch to at most 8 top-level nodes or approximately 30,000 compacted characters.
• Stop deepening a screen after its user-visible copy, states, actions, and data needs are understood.
• Deduplicate text while preserving first-seen order. Normalize whitespace but do not rewrite product copy.
• Summarize colors, typography, and spacing as tokens/patterns; do not repeat identical style objects per child.
• Keep raw responses out of conversation output. If a raw response must be processed outside the tool call, pass it through scripts/compact_figma_tree.py.

Evidence hierarchy

Use this precedence unless the user states otherwise:

1. Explicit user decisions
2. Approved meeting/product notes
3. Current Figma behavior and copy
4. Existing production contracts
5. Current implementation details

Distinguish confirmed requirements, reasonable inferences, and unresolved decisions.

References

• Read references/figma-bridge.md when choosing bridge calls, batching, or screenshots.
• Read references/output-schemas.md when producing a feature brief, state model, contract matrix, or gap audit.