penpot-build-screen

penpot-build-screen — brief to on-system screen

1. Title + How it works

penpot-build-screen turns a brief into a crafted, on-system screen; every mutation goes through execute_code; validate visually with export_shape; read structure with penpotUtils.shapeStructure (full tool surface: shared/penpot-mcp-tool-reference.md). It first discovers the existing design system (tokens + components via penpot-foundations / the local library), then builds the screen section by section inside flex Boards, binding tokens and reusing components — never generating an entire screen in one call.

2. The One Rule That Matters Most

Reuse the system; build incrementally. Prefer existing components and semantic tokens over raw shapes. Build one section, checkpoint with an export_shape, then continue. If no system exists, bootstrap only a minimal one (hand off to penpot-foundations for anything substantial).

3. Penpot MCP Tool Reference

Full surface: shared/penpot-mcp-tool-reference.md. Key calls: penpot.createBoard()/addFlexLayout() for the screen and sections; penpot.library.local.components + comp.instance() to reuse components; shape.applyToken for token binding; export_shape('selection'|'page') at checkpoints.

4. Plugin API Essentials

Gotcha numbers refer to shared/plugin-api-gotchas.md.

• Screen and sections are Boards with flex layout; compose with append order + gaps + align/justify.
• Reuse components via component.instance(); don't redraw system parts as raw shapes.
• #3 text auto-sizing — set growType to auto-width/auto-height AFTER any resize() (which forces fixed).
• #4 flex overrides child x/y — use layoutChild for stretch/margins.
• Bind colors/spacing/radius/type to semantic tokens, never hardcoded.
• #11 every new Board is born with an OPAQUE WHITE fill — this skill's critical failure mode. createBoard() ships fills = [{ fillColor: "#FFFFFF", fillOpacity: 1 }]; left in place, a structural wrapper's square white corners poke out behind rounded children (radius looks broken), and layout boards keeping the literal white never flip in dark mode (no token, off-system). Fill policy: decide per board — a structural board (layout-only chrome: sections, rows, wrappers) gets board.fills = []; a surface (screen bg, card, sheet, button) binds a color.bg.* token, never a literal. Default layout containers to transparent (shared/modes-and-policies.md).
• #13b exact font match — penpot.fonts.all.find(f => f.name === "…"); findByName is a substring search.
• #15 — set layoutChild.* only AFTER appendChild; openPage is async (two-call protocol: create+open, then assert currentPage.id before any createBoard).
• #16 — run clearFlip after every section (fill-sized children can come back with flipX = true; canonical helper in shared/visual-effects.md).
• #17 — await board.waitForLayoutUpdate() before reading geometry (replaces the sleep-100ms idiom).
• #18 — image fills: const img = await penpot.uploadMediaUrl(name, url); shape.fills = [{ fillOpacity: 1, fillImage: img }], verified in the next call — see shared/visual-effects.md.
• Verify unfamiliar signatures with penpot_api_info first.

5. Token-Aware Brief Contract

• Context — product, audience, platform, brand mood.
• Objective — the single screen and its primary user goal.
• Inputs — content/sections, existing tokens & components, style profile, viewport size.
• Constraints — use existing components; on-grid spacing; semantic tokens only; forbidden patterns; none of the named tells in shared/design-quality.md §7.
• Acceptance Criteria — clear hierarchy; 4px rhythm; AA contrast; reuses system; responsive intent stated; design-quality score ≥ 3 on all seven axes (shared/design-quality.md §8) or the weak axes explicitly presented.

Act as a senior product/visual designer who makes deliberate aesthetic decisions, not generic ones.

6. Mandatory Workflow

Visual self-review (mandatory): before every ✋ checkpoint that shows visual work, run the export → look → fix loop from shared/visual-self-review.md — export the unit you just built, inspect the image yourself against the checklist, fix visible defects (max 2 iterations), and present that same export with any remaining defects named.

Phase 0 — Discovery. high_level_overview; inventory tokens/components (scripts/setupOrReuseSystem.js); analyze the brief (references/01-brief-analysis.md); pick a style profile (references/02-style-profiles.md) and a named screen skeleton (shared/design-quality.md §5) — check the ledger for prior screens this session and apply the variety rule (differ on ≥ 1 axis, say which). ✋ Checkpoint: confirm brief + style + skeleton + section list.

Phase 1 — Frame. Create the screen Board with scripts/createScreenFrame.js (idempotent; returns ids; binds bg/gap/padding tokens). Set viewport size. ✋ Checkpoint.

Phase 2..N — Sections. Build each section as a tokenized flex Board reusing components (scripts/buildSection.js). One section per execute_code call. ✋ Checkpoint after each (export_shape).

Phase N+1 — Assemble & critique. scripts/assembleScreen.js composes sections; scripts/auditScreenQuality.js checks layout coverage (every board has flex/grid), token binding, on-grid spacing, naming. A non-empty boardsWithoutLayout fails the gate — add a layout to each flagged board before reporting done. Then run the scored critique (references/05-critique-framework.md): score the final export 1–5 on the seven axes of shared/design-quality.md §8; any axis < 3 → targeted revision (max 2 passes), then emit the design-quality report (Markdown + JSON per shared/report-schemas/design-quality-report.schema.json, mirrored to the ledger). Report.

7. Critical Rules

1. Flex by default — every container is a layout Board. The instant you create a Board, give it a

flex (addFlexLayout()) or grid (addGridLayout()) layout before appending children — the screen root, every section, AND every nested grouping (card, row, list, stat cluster, form field, button group). NEVER arrange UI elements with absolute x/y, and NEVER use a plain Group to lay out UI. A Board without a layout is a bug; the Phase N+1 audit (boardsWithoutLayout) fails the build if any exist.

1. Reuse existing components before creating raw shapes.
2. Bind to semantic tokens; never hardcode color/spacing/radius/type.
3. One section per call; checkpoint with export_shape.
4. All spacing on the 4px grid; consistent rhythm.
5. Don't bootstrap a full design system here — hand off to penpot-foundations.
6. State responsive intent (sizing: fill/auto/fix) explicitly.

8. Domain Architecture

A screen = a root flex Board (column) → section Boards (each its own flex) → component instances + tokenized primitives. Composition follows a clear visual hierarchy: primary action prominent, content grouped, whitespace on the spacing scale, type from the semantic type tokens.

9. Modes & Policies

Default review. Geometry/layout changes always require a checkpoint (shared/modes-and-policies.md).

10. State Management

Ledger under RUN_ID: phase, screenBoardId, sections:[{name,id,done}], styleProfile, skeleton, designQuality (the §8 report object). Resume by re-reading structure.

11. User Checkpoints

12. Naming Conventions

shared/naming-conventions.md: layers semantic HTML (header, main, section, nav, button); screen Board named for the view (Dashboard).

13. Anti-Rationalization Table

14. Helper Code Snippets

// Screen frame (Phase 1)
const screen = penpot.createBoard();
screen.name = "Dashboard";
screen.resize(1440, 1024);
const flex = screen.addFlexLayout();
flex.dir = "column";
flex.horizontalSizing = "fix"; flex.verticalSizing = "auto";
penpot.currentPage.root.appendChild(screen);         // append first (#7); layoutChild only after append (#15)
// SPACING: bind tokens, never literals (#8: padding tokens bind on ≥ 2.17; numeric mirror + ledger exception otherwise)
const gap = penpotUtils.findTokenByName("spacing.24");
if (gap) screen.applyToken(gap, ["rowGap"]); else flex.rowGap = 24;
const pad = penpotUtils.findTokenByName("spacing.32");
const FLEX_PROP = { paddingTop: "topPadding", paddingBottom: "bottomPadding", paddingLeft: "leftPadding", paddingRight: "rightPadding" };
for (const side of Object.keys(FLEX_PROP)) {
  try { if (!pad) throw new Error("no token"); screen.applyToken(pad, [side]); }
  catch (e) { flex[FLEX_PROP[side]] = pad ? Number(pad.resolvedValue) : 32; }   // record { kind: "padding-mirrors-token" } in the ledger
}
// FILL POLICY (gotchas #11): the screen root is THE one surface — bind its bg to a token so it flips in
// dark mode. Every section nested inside stays transparent (buildSection.js clears their default white).
screen.fills = [];                                   // drop Penpot's default opaque #FFFFFF
const bg = penpotUtils.findTokenByName("color.bg.default");
if (bg) screen.applyToken(bg, ["fill"]);             // bound surface, not a literal
storage.bs = { screenBoardId: screen.id };
return { screen: screen.id, bgBound: !!bg, gapBound: !!gap, padBound: !!pad };

15. Reference Resources

• penpot_api_info('Board'), penpot_api_info('LibraryComponent'), penpot_api_info('Text').

16. Supporting Files

references/: 01-brief-analysis.md, 02-style-profiles.md, 03-layout-composition.md, 04-component-recipes.md, 05-critique-framework.md, 06-anti-rationalization.md, 07-error-recovery.md; effects, image placeholders, gradients: shared/visual-effects.md. scripts/: setupOrReuseSystem.js, createScreenFrame.js (Phase 1 frame: idempotent, token-bound bg/gap/padding, ledger), buildSection.js, assembleScreen.js, auditScreenQuality.js.

Doctrine paths. shared/… and policies/… resolve inside this bundle in native installs (vendored by the installer); in a Claude Code plugin install they live at the plugin root — ${CLAUDE_PLUGIN_ROOT}/shared/…, two directories up from this file.