| name | generating-figma-design |
| description | Use when the user requests to design, create, build, generate, or make a new Figma component or screen — including phrases like "make a button", "design a settings page", "build a new card", "generate X". Produces a CSpec, compiles it to a scene graph, executes it in Figma via MCP, and verifies the output. |
{{ACTIVE_RULES}}
Generating Figma Design
Overview
Unified make flow: CSpec → scene graph → compiler → Figma execute → screenshot → verify. Replaces the legacy spec → design → review cycle with a single compiler-driven action. All tokens resolve against the knowledge base; all Figma API rules are enforced by the compiler.
When to Use
Invoke when the user:
- says
make, design, create, build, generate, or asks for a "new component" / "new screen"
- has already run
setup (knowledge base exists)
- has an MCP transport available (console or official)
Do NOT use if:
- the user is adjusting an existing design that is already in Figma — use
learning-from-corrections instead
- the user is archiving or shipping — use
shipping-and-archiving
- the knowledge base is missing — run
extracting-design-system first
Procedure
Before starting, load:
references/compiler-reference.md (repo-root) — scene graph JSON format and rules
references/transport-adapter.md (repo-root) — transport detection and tool mapping
Prerequisites
- Knowledge base exists (registries populated) — if not: "Run
setup first"
- MCP transport available (see
references/transport-adapter.md (repo-root) Section A)
Phase A — Context (target: 30s)
A1. Detect transport
Read references/transport-adapter.md (repo-root) Section A. Determine console vs official transport.
Console: figma_get_status() -> setup.valid: true
Official: whoami() + test use_figma call
Report:
Transport: {console | official}
A2. Load registry index
Load the summary of available DS components — names and types only, not full registry data:
registries/components.json — extract component names, variant property names, and keys
registries/variables.json — extract variable name paths (for token reference validation)
registries/text-styles.json — extract style names (for $text/ reference validation)
registries/icons.json — extract icon names (if file exists)
registries/logos.json — extract logo names (if file exists)
Do NOT load guides, patterns, or figma-api-rules.md. The compiler handles all Figma API rules.
A2b. Check KB freshness (drift guard)
The compiler guarantees every $token resolves against the KB — but NOT that the
KB still matches the live Figma library. A token removed/renamed in Figma but
still present in a stale KB will compile fine and then fail at execute time.
This check is a runtime guard; it deliberately lives here (not in the compiler),
because a time-based check would break the compiler's determinism law.
Read generatedAt from each loaded registry and compare to today's date:
- ≤ 7 days → fresh, no warning.
- > 7 days → surface: "⚠ KB last synced {N}d ago — the refresh cron may not be
running; tokens/components may have drifted from live Figma."
- > 30 days, or
generatedAt missing → STRONGLY warn and recommend re-running
setup (or the KB cron) before generating: resolved tokens that were removed or
renamed in Figma will pass compile but fail on execute.
Carry the KB age into the C4 plan ("KB age: {N}d"). This does not block generation
— it informs the user before they commit to a make.
A3. Load learnings
Load knowledge-base/learnings.json (skip if file doesn't exist).
Filter by context matching the user's description:
- Include all global learnings (
scope: "global")
- Include contextual learnings where
context.screenType or context.component matches
A4. Load recipe index
Load knowledge-base/recipes/_index.json (skip if file doesn't exist — no recipes yet).
Phase B — Recipe Match
B1. Extract archetype
From the user's description, identify:
- Mode: component or screen (ask if ambiguous)
- Archetype: screen type (settings, dashboard, form, detail, list...) or component type
- Keywords: key terms from the description (sidebar, form, table, cards, navigation...)
B2. Score against recipe index
For each recipe in _index.json, compute a match score:
| Dimension | Weight | Method |
|---|
| Archetype match | 0.40 | Exact match on meta.archetype vs extracted archetype |
| Tag overlap | 0.25 | Jaccard similarity between recipe tags and extracted keywords |
| Structural match | 0.20 | Zone count, component types, parameter compatibility |
| Confidence | 0.15 | Recipe's current confidence score |
B3. Apply match result
| Score | Action |
|---|
| >= 0.85 | Exact match. Load recipe file, pre-fill CSpec from recipe parameters. Report: "Recipe match: {name} (score: {score}). Using as template." |
| 0.60 -- 0.84 | Partial match. Load recipe as scaffold. Report: "Partial recipe match: {name} (score: {score}). Using as starting point, will supplement missing zones." |
| < 0.60 | No match. Proceed from scratch. Report: "No recipe match. Generating from scratch." |
Phase C — CSpec (target: 30-60s)
C0. Decompose into sections, then classify each (design intelligence)
Before writing any node, break the request into its major sections, then decide
HOW each section is built. This is where DS fidelity and composition are won.
1. Decompose the request into major sections, top to bottom:
- Screen → e.g. Header, Hero, Content panels, Footer.
- Modal / dialog → Title bar, Body / form sections, Action bar.
- Drawer / sidebar / panel → Navigation, Content area, Footer actions.
- Component → its part structure (container, leading, label, trailing, states).
2. Classify each section into exactly one bucket — this decides INSTANCE vs build:
exact — a published DS component (or one of its variants) covers the whole
section → one INSTANCE node; set its variant + properties. Preferred.
compose — no single component fits, but the section is built from DS pieces
(component INSTANCEs inside tokenized FRAMEs) → never raw shapes with hardcoded
values.
new — no DS component covers it and it is not composable from existing ones
→ hand to C3 (new_components); the component is built first.
Default preference: exact > compose > new. A raw RECTANGLE/ELLIPSE whose
name matches a DS component is a red flag — the compiler warns (Rule 18); use an
INSTANCE instead.
Report the section map before C1:
SECTIONS:
{Section} → {exact: Component(variant) | compose | new: name}
C1. Generate CSpec YAML
Choose the appropriate template:
- Screen mode:
skills/generating-figma-design/references/templates/screen-cspec.yaml
- Component mode:
skills/generating-figma-design/references/templates/component-cspec.yaml
Fill the CSpec based on:
- User description (intent, sections, components)
- Recipe template (if match found in Phase B)
- Registry data (available components, tokens, text styles)
If a recipe was matched (>= 0.60):
- Start from the recipe's
graph structure
- Replace
{{ param }} placeholders with values from the user's description
- Resolve
@lookup:ComponentName references against the live registry
- Add/remove zones as needed for the specific request
If no recipe match (from scratch):
- Build the layout tree node by node using the CSpec template structure
- Reference DS components as INSTANCE nodes (by name — the compiler resolves keys)
- Use
$token references for all spacing, colors, radius, typography — and pick
the most specific token that carries intent. When collections alias each
other (primitive ← semantic ← component), reference the most specific one:
prefer a semantic token ($color/bg/...) over a raw palette value, and a
component-level token over a semantic one when it exists. Never reference a
primitive when a semantic/component token expresses the same intent — a
primitive ignores theme/mode switches.
- Choose tokens by intent, not by resolved value. A token's value depends on
the active mode (light/dark); selecting by the value you see in the default
mode silently breaks the other modes.
- Use REPEAT nodes for lists/grids with repeated structure
C2. Apply learnings
Integrate learnings from A3 into the CSpec:
- Global learnings (
scope: "global"): auto-apply — replace default token values with learned preferences
- Contextual learnings (
scope: "contextual"): suggest — note in comments, apply if context matches
Report applied learnings:
LEARNINGS APPLIED ({n}):
- {rule} (signals: {n}, scope: {scope})
- {rule} (signals: {n}, scope: {scope})
C3. Detect new DS components (screen mode only)
Take the sections C0 classified as new (the patterns no existing DS component
covers and that are not composable from existing ones). Add each to the
new_components section of the CSpec. Sections classified exact become INSTANCE
nodes; sections classified compose become FRAMEs of INSTANCEs — neither reaches
this step.
If new components are identified:
{N} new DS component(s) needed:
1. {name} — {description}
These must be created before this screen. Starting with: {name}
-> Trigger a nested make flow for each new component. When all are done, resume the screen make.
C4. Present plan to user
Show a readable summary of the CSpec (NOT raw YAML). Format as a plan tree:
PLAN: {name}
Mode: {screen | component}
Canvas: {1440px (web) | 390px (mobile) | 1024px (tablet)}
Recipe: {recipe name or "from scratch"}
Learnings: {n} applied
KB age: {N}d{ ⚠ stale — consider `setup` if > 30d}
STRUCTURE:
Root ({width}x{height}, {layout direction})
+-- {Zone 1} ({width}, {layout})
| +-- {Component} ({variant})
| +-- {Component} ({variant})
+-- {Zone 2} ({fillH}, {layout})
| +-- {Section 1}
| | +-- {Component} ({variant})
| +-- {Section 2}
| | +-- REPEAT x{n}: {Component}
DS COMPONENTS: {n} instances
TOKENS: {n} references
STATES: {list or "populated only"}
Generate this design?
C5. User validates
Wait for explicit user confirmation. The user can:
- Approve -> proceed to Phase D
- Adjust -> modify the CSpec based on feedback, re-present plan
- Cancel -> abort
C6. Save CSpec
Save the CSpec YAML to specs/active/{name}.cspec.yaml.
Phase D — Compile + Execute
D1. Convert CSpec to scene graph JSON
Transform the CSpec's layout tree into the scene graph JSON format defined in references/compiler-reference.md (repo-root):
- CSpec
layout nodes map directly to scene graph nodes
- All
$token references are preserved (the compiler resolves them)
- Component names in INSTANCE nodes are preserved (the compiler resolves keys)
- REPEAT and CONDITIONAL nodes pass through to the compiler
Add the root wrapper:
{
"version": "3.0",
"metadata": {
"name": "{CSpec meta.name}",
"width": {CSpec meta.width},
"height": {CSpec meta.height},
"transport": "{detected transport}",
"fileKey": "{user's file key}"
},
"fonts": [ ... ],
"nodes": [ ... ]
}
Font list: Collect all unique font families and styles referenced by $text/ tokens in the scene graph. Cross-reference against registries/text-styles.json to get actual font family + style values.
D2. Write scene graph to temp file
cat > /tmp/bridge-scene-{name}.json << 'EOF'
{ ... scene graph JSON ... }
EOF
D3. Run the compiler
bridge-ds compile \
--input /tmp/bridge-scene-{name}.json \
--kb {kb-path} \
--transport {console|official}
The compiler outputs a JSON array of { id, code, description } chunks to stdout.
D4. Handle compiler errors
If the compiler returns errors to stderr:
- Read the error messages and suggestions
- Fix the scene graph JSON based on the suggestions (see
references/compiler-reference.md (repo-root) Section 8 for common errors)
- Re-write the temp file and re-run the compiler
- Maximum 3 attempts. If still failing after 3, report errors to user and ask for guidance.
D5. Execute chunks in Figma
For each output chunk from the compiler:
Console transport:
figma_execute({ code: "{chunk.code}" })
Official transport:
use_figma({
fileKey: "{fileKey}",
description: "{chunk.description}",
code: "{chunk.code}"
})
Execute chunks sequentially. If a chunk fails:
- Read the error
- Report to user
- Attempt fix if the error is clear (e.g., font not loaded, component not found)
D6. Take screenshot
Take a screenshot AFTER the final chunk (not after each chunk):
Console: figma_take_screenshot({ node_id: "{rootNodeId}", file_key: "{fileKey}" })
Official: get_screenshot({ nodeId: "{rootNodeId}", fileKey: "{fileKey}" })
D7. Save snapshot
Save a snapshot of the design's node tree for future fix diffing.
Run a node tree extraction script via figma_execute (or use_figma), using the root node ID from D5.
Save to specs/active/{name}-snapshot.json:
{
"meta": {
"spec": "{name}",
"generatedAt": "{ISO timestamp}",
"rootNodeId": "{rootId}",
"fileKey": "{fileKey}",
"recipe": "{recipe ID or null}",
"learningsApplied": ["{learning IDs}"]
},
"tree": { ... extracted node tree ... }
}
Phase E — Present
E1. Show screenshot
Display the screenshot taken in D6.
E2. Report
Design compiled and executed.
File: {figma_url}
Created:
- {n} component instances
- {n} bound variables (colors + spacing + radius)
- {n} learnings applied
- Recipe: {recipe name or "from scratch"}
- Chunks: {n} executed
Warnings:
- {any issues}
E3. Offer next step
Looks good? Options:
- Describe changes -> I'll modify and recompile
- "I adjusted in Figma" -> triggers fix flow
- "done" / "ship it" -> triggers done flow
Iteration Loop
User describes changes (in conversation)
- Identify which parts of the scene graph need modification
- Update the scene graph JSON (modify affected nodes only)
- Re-write temp file
- Re-run the compiler
- Execute only the affected chunks (not the entire design)
- Take new screenshot
- Update snapshot
- Present result and offer next step again
User says "I adjusted in Figma"
Trigger the fix flow via the learning-from-corrections skill.
User says "done" / "ship it"
Trigger the done flow via the shipping-and-archiving skill.
Target Turn Budget
| Turn | Action |
|---|
| 1 | Phase A (context load) + Phase B (recipe match) |
| 2 | Phase C (CSpec generation, present plan to user) |
| 3 | User confirms or adjusts |
| 4 | Phase D (compile + execute all chunks) |
| 5 | Phase E (screenshot, report, offer next step) |
| 6+ | Iteration loop (if changes requested) |
Target: 5 turns for a first-pass generation. Each iteration adds 2-3 turns.
Compiler vs Raw Scripts
In v3, Claude NEVER writes raw Figma Plugin API scripts. The workflow is:
Claude produces scene graph JSON
-> Compiler resolves tokens, validates structure, generates code
-> Code chunks are executed via MCP
NOT:
Claude writes figma_execute scripts directly
This means:
- No need to load
figma-api-rules.md
- No need to remember FILL-after-appendChild, resize-before-sizing, etc.
- No pre-script element audits (the compiler validates against registries)
- No script line counting or splitting logic
- Errors are caught by the compiler with actionable suggestions
The only time Claude touches raw Plugin API code is:
- Snapshot extraction scripts (small, standardized)
- Emergency fixes when the compiler cannot handle an edge case (rare)
Every Figma execution MUST pass Gate A (compile) before execute and
Gate B (visual) before claiming the iteration is complete. See
`references/verification-gates.md` (repo-root).
NEVER write raw Figma Plugin API code. All scene graphs go through
lib/compiler/compile.ts (shipped as bridge-ds compile).
NEVER use hardcoded primitive values in the scene graph. Only
$token references.
Red Flags
See the full catalog at references/red-flags-catalog.md (repo-root).
Top flags for this skill:
- "I'll hardcode this hex once, it's faster" → Always use a semantic token.
- "The compiler is overkill for this tiny thing" → The compiler is the only path.
- "I remember this nodeId from my last session" → NodeIds are session-scoped, re-search.
Verification
This skill is gated by references/verification-gates.md (repo-root):
- Gate A (Compile) — mandatory before any
figma_execute / use_figma.
- Gate B (Visual) — mandatory at the end of each iteration. Fresh screenshot in this turn + user confirmation.
Evidence to surface: compiler stdout, scene graph JSON, screenshot tool result, user confirmation text.
Skill-specific references
./references/templates/component-cspec.yaml — CSpec template for components
./references/templates/screen-cspec.yaml — CSpec template for screens
The make flow (decision diagram)
digraph make_flow {
"User says 'make X'" [shape=doublecircle];
"Load context (recipes, learnings)" [shape=box];
"Generate CSpec" [shape=box];
"Compile" [shape=box];
"Compile exit 0?" [shape=diamond];
"Surface compile error" [shape=box style=filled fillcolor=lightcoral];
"Execute via MCP" [shape=box];
"Screenshot" [shape=box];
"User satisfied?" [shape=diamond];
"Gate A passed" [shape=doublecircle style=filled fillcolor=lightgreen];
"Capture intent diff" [shape=box];
"User says 'make X'" -> "Load context (recipes, learnings)";
"Load context (recipes, learnings)" -> "Generate CSpec";
"Generate CSpec" -> "Compile";
"Compile" -> "Compile exit 0?";
"Compile exit 0?" -> "Surface compile error" [label="no"];
"Surface compile error" -> "Generate CSpec";
"Compile exit 0?" -> "Execute via MCP" [label="yes"];
"Execute via MCP" -> "Screenshot";
"Screenshot" -> "User satisfied?";
"User satisfied?" -> "Gate A passed" [label="yes"];
"User satisfied?" -> "Capture intent diff" [label="no"];
"Capture intent diff" -> "Generate CSpec";
}