Inspect before editing. Verify the Paper file, selected node, GPUI version, existing components, theme, assets, and dirty worktree first.
Use Paper MCP evidence, not a screenshot alone. Capture hierarchy, computed styles, text, fonts, assets, and a 2x screenshot of the exact target.
Treat Paper JSX as a structural hint, never as GPUI code or unquestioned truth. Computed styles and the screenshot settle ambiguity.
Preserve the app shell, state model, component conventions, and platform behavior. Replace only the visual surface in scope.
Prefer GPUI layout over coordinate transcription. Use flex/grid for normal structure and absolute positioning only where the design genuinely overlaps.
Do not rasterize text, controls, panels, or whole screens to fake fidelity. Export only real visual assets such as icons, illustrations, textures, and photos.
Validate the native runtime. cargo check is necessary but does not prove fonts, window chrome, scale factor, focus, hover, clipping, or pixel fidelity.
Work in small regions. Large artboards must be translated and compared section by section before the whole screen is judged.

Workflow

1. Establish both contexts

Run the read-only project inspector:

scripts/inspect_gpui_project.sh /path/to/gpui-project

Then inspect the checkout directly:

Confirm the repository root, branch, dirty state, and requested surface.
Read the relevant Cargo.toml, lockfile entry, app entrypoint, root view, theme/tokens, reusable components, asset source, and nearby tests.
Identify whether the project uses published gpui, a Git revision, a workspace checkout, a fork, or a wrapper component library.
Follow the pinned checkout's APIs when they differ from examples. GPUI is pre-1.0 and changes frequently.

Use the Paper MCP server:

Call get_basic_info to verify the currently open file and artboards.
Call get_selection to resolve the target.
If the selection is empty or includes unrelated nodes, ask the user to select one artboard or frame. Do not guess from names when multiple targets fit.
Default to read-only Paper tools. Do not change the design unless the user explicitly asks for design edits.

Read paper-mcp.md before extracting a non-trivial design or troubleshooting Paper connectivity.

2. Capture a design evidence pack

Acquire evidence in this order:

get_screenshot at 2x for the selected root.
get_node_info and get_tree_summary for dimensions and hierarchy.
get_jsx for a compact structural interpretation.
get_computed_styles in batches for the root, layout containers, text, controls, separators, and visually distinct descendants.
get_font_family_info for every non-system family and used weight/style.
get_fill_image or export for real image/vector assets.
get_children plus targeted subtree calls when a large design exceeds tool limits or loses detail.

Record a compact evidence table before coding:

Paper node	Role	Bounds	Layout	Spacing	Type	Paint/effects	Asset	Behavior

Include exact values where they affect fidelity. Do not fill unknowns with plausible defaults. Re-query the node or label the uncertainty.

3. Build a translation plan

Map the design into these layers:

Window and chrome — viewport, titlebar, background, safe inset, minimum size, and platform-specific frame.
Structural regions — sidebar, toolbar, content columns, inspector, footer, modal, or overlay.
Reusable primitives — button, icon button, field, row, badge, separator, list item, empty state, and section heading.
Tokens — color, spacing, typography, radius, border, shadow, and breakpoint constants.
State and behavior — selection, hover, pressed, focus, disabled, loading, scrolling, shortcuts, and resize behavior.

Reuse existing tokens and components when their rendered result matches. Extend them narrowly when they do not. Do not create a parallel design system for one screen.

Read gpui-translation.md before implementing. It contains the detailed Paper/CSS-to-GPUI mapping and the boundaries where GPUI needs a custom element, canvas, or deliberate approximation.

4. Implement from geometry inward

Use this order because it minimizes rework:

Match window/content bounds and large background regions.
Match flex direction, fixed/flexible dimensions, gaps, padding, alignment, wrapping, overflow, and clipping.
Match typography: actual family, available weight, size, line height, wrapping width, alignment, truncation, and baseline.
Match fills, borders, radii, opacity, shadows, gradients, and separators.
Add exported assets at their intended logical size.
Add interaction states and focus behavior.
Extract reusable GPUI components only after the repeated visual pattern is confirmed.

Keep exact Paper pixels as px(...) during the first fidelity pass. Consolidate repeated values into project tokens after the screen matches. Avoid premature rounding to GPUI convenience scales such as .gap_3() when Paper specifies an off-scale value.

5. Validate behavior and visual fidelity

Run the nearest repository checks first:

cargo fmt --check
cargo check -p <owning-crate>
cargo test -p <owning-crate>
cargo clippy -p <owning-crate> --all-targets -- -D warnings

Adapt commands to the repository. Do not claim checks that were unavailable or unreasonably broad.

Then launch the real app and compare:

Reproduce the Paper artboard's logical viewport inside the GPUI content area.
Confirm OS scale factor, theme, font availability, and window chrome.
Put the app in the same state and use the same content as Paper.
Capture the GPUI window or content region without resizing the result.
Compare side by side, overlay at partial opacity, and use a difference image when tooling is available.
Fix mismatches in this order: bounds, layout, typography, paint, assets, interaction polish.
Repeat until further changes are below the agreed tolerance.

Read fidelity-validation.md for capture normalization, diff techniques, tolerances, and mismatch diagnosis.

6. Prove completion

Do not call the work complete until:

The exact Paper file and target node are identified.
The relevant GPUI crate builds and targeted tests pass.
The native app has been launched at least once.
A final Paper screenshot and GPUI screenshot exist at matching logical bounds.
Major structure, typography, colors, radii, and assets have been visually compared.
Hover, active, focus, scrolling, resizing, and keyboard behavior relevant to the screen have been exercised.
Remaining deltas and platform limitations are stated plainly.

Reference routing

Read paper-mcp.md for connection setup, safe tool sequencing, node extraction, tokens, fonts, assets, and Paper failure modes.
Read gpui-translation.md for current GPUI architecture, styling APIs, property mapping, components, assets, state, and version shields.
Read fidelity-validation.md for the screenshot loop, visual tolerances, diff workflow, responsive checks, and native acceptance.
Read worked-example.md when starting a new translation or when Paper's JSX/CSS structure does not map cleanly to GPUI.

Failure shields

If Paper tools are absent, stop and explain how to connect Paper Desktop MCP. Do not reconstruct a design from memory.
If Paper reports the wrong file, have the user open the intended file and call get_basic_info again.
If no single target is selected, ask for a selection or an exact node ID.
If a Paper subtree is huge, split by structural region and keep one screenshot of the full target for global alignment.
If an exact font is unavailable to GPUI, do not silently substitute it. Report the missing family/weight, add or register the font if authorized, then recapture.
If a Paper effect has no direct GPUI equivalent, preserve hierarchy and interaction, implement the nearest maintainable native effect, and document the delta. Consider canvas or a custom Element only after ordinary GPUI styling is proven insufficient.
If screenshots differ despite equal CSS-like values, inspect content bounds, device scale, text metrics, default line height, border inclusion, and OS window chrome before nudging arbitrary pixels.

Final response

Report:

Paper file, artboard/frame, and viewport used.
GPUI files and reusable components changed.
Assets and fonts added or reused.
Build, test, launch, and screenshot checks performed.
Remaining visual or behavioral deltas.
Any assumptions that still require user confirmation.

paper-to-gpui

Paper to GPUI

Non-negotiable rules