/build-gpui-apps

Forward-test scenarios

Run these after substantial changes to the umbrella skill. Give only the prompt block to a fresh agent with no creator context. Review the result against the signals separately; do not leak the rubric into the prompt.

1. Accessible cancellable search

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. Read-only: inspect the pinned
GPUI testing example and propose a focused implementation plan for a
cancellable, accessible GPUI search panel with keyboard actions and a
virtualized result list. Do not edit files. Report the exact reference layers
you used, concrete state/action/task ownership, tests, and any pinned-version
uncertainties.

Review signals:

• treats the target checkout as API authority and respects its selected stack;
• routes through architecture, async/performance, accessibility, and testing;
• gives the view/model one clear source of truth;
• holds or replaces the search task and rejects stale generations;
• uses typed actions and scoped key contexts;
• virtualizes results with stable semantic IDs;
• covers focus, announcements, loading/empty/error, cancellation, and a real launch path;
• distinguishes compiler/test evidence from runtime proof.

2. Honest Apple-style toolbar

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. Read-only: inspect a pinned GPUI
hello-world example and propose how to add an Apple-polished floating toolbar
on macOS while preserving honest Linux and Windows behavior. Do not edit
files. State the material capability tier, fallbacks, architecture,
interactions, accessibility, and validation evidence you would require.

Review signals:

• inspects the target's real startup, theme, and platform policy;
• names native Liquid Glass only behind an actual guarded AppKit capability;
• does not call translucent GPUI paint native backdrop blur;
• keeps content mostly solid and glass limited to controls/navigation;
• provides opaque, high-contrast, reduced-transparency, inactive-window, and cross-platform fallbacks;
• covers focus, keyboard, pointer, reduced motion, scale, resize, cleanup, and screenshot evidence;
• avoids speculative native bridging when the target does not need it.

3. Paper connection unavailable

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. I have a selected Paper.design
frame and want it translated into a named GPUI fixture. Read-only: inspect what
is locally available and give the next concrete steps; do not edit files. Be
explicit about what evidence you can and cannot obtain and what must happen
before implementation.

Review signals:

• inspects the GPUI target without modifying it;
• requires a live Paper MCP connection and exact selected frame/node;
• does not infer geometry, styles, assets, or fonts from memory;
• explains the minimal connection/selection recovery path;
• lists screenshot, hierarchy, computed-style, token, font, and asset evidence to capture after connection;
• stops before implementation and keeps the unavailable path explicit.

4. Unicode and IME-safe editor review

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. Read-only: inspect the pinned
GPUI text-input example and review a proposed custom single-line editor that
stores selection offsets as usize, replaces text on every marked-text update,
and hit-tests with average glyph width. Explain the production design and test
plan; do not edit files.

Review signals:

• identifies the ambiguous offset unit as unsafe;
• separates Rust UTF-8 bytes, platform UTF-16 code units, grapheme boundaries, and shaped visual positions;
• treats marked text as provisional composition;
• covers the full pinned EntityInputHandler contract;
• derives candidate bounds and hit tests from current shaped layout;
• centralizes edit transactions, selection, undo, notification, and revision checks;
• includes emoji, combining marks, bidirectional text, real IME, scroll/resize, clipboard, focus, accessibility, and stale-layout tests;
• recommends an existing maintained editor first when appropriate.

5. Production starter from the example repository

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. Set up a production-ready GPUI
starter for a new desktop product named Northstar using
https://github.com/lassejlv/gpui-starter as the example. The target directory
already exists and contains an unrelated README. Work read-only: inspect the
source/example and return an implementation plan, proposed project tree,
identity ledger, and exact acceptance gates. Do not edit, clone over, delete,
or publish anything.

Review signals:

• protects the existing non-empty target and does not delete or overwrite it;
• inspects and records the current example commit, lockfile GPUI revision, and matching toolchain rather than trusting floating main;
• asks for or marks provisional the owned reverse-DNS app ID, publisher, platforms, minimum OS versions, distribution, data, and update owner;
• treats gpui-starter as a minimal two-crate baseline, not a production-ready artifact by itself;
• preserves the desktop/ui split and adds a headless crate only for proven domain/service ownership;
• covers full identity rename, observable startup, configuration/migrations, secrets, diagnostics, accessibility, lifecycle-owned async work, and one real vertical slice without adding irrelevant subsystems;
• establishes the framework choice; pins the Rust toolchain and selected Kit release/matching GPUI snapshot or upstream Git revision, commits the lockfile, and uses --locked CI with immutable action pins and least privilege;
• distinguishes compile/test, real launch, installed artifact, signing, and upgrade evidence per claimed platform;
• identifies whole-window blur as a capability/tradeoff, includes an opaque fallback, and does not call it per-control Liquid Glass;
• gives macOS, Windows, and Linux packaging only for platforms actually selected, including signing/identity/install/upgrade checks;
• reports unresolved production facts as release blockers rather than silently inventing them.

6. Recommend Kit, ask before adoption

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. I want a new native Rust desktop
notes app. No framework choice has been made. This is a read-only dry run:
return the next user-facing response and what work may proceed before an
answer. Do not edit files, run installation commands, or ask through a tool.

Review signals:

• recommends Kit with concrete reasons and asks Kit versus upstream GPUI;
• waits for an answer before dependency/bootstrap implementation;
• continues read-only preparation rather than inventing consent;
• does not ask whether the skill update itself may proceed.

7. Kit already selected: imports, state, overlays, design, testing

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. We have already chosen GPUI Kit
0.7.0. Plan a settings window with a retained text input, a controlled switch,
and an About dialog. Work read-only. Show the dependency/import roots and
bootstrap shape, explain state/subscription and overlay ownership, and cite
the bundled design, coding, component, and UI integration testing references.
Do not ask the framework question again or claim you compiled the plan.

Review signals:

• uses gpui-kit and gpui_kit; avoids separate direct GPUI/Component/Base deps;
• reads actual normative Design/Coding Guides rather than their summaries;
• uses current Kit open_window, returns content, lets Root render overlays;
• retains input state/subscriptions, uses on_change for the controlled switch;
• uses tokens, relevant extension traits, stable IDs, focus, and error states;
• identifies test-support and #[gpui_kit::test], respecting feature availability;
• does not use the raw fixture as Kit compilation evidence.

8. Preserve an explicit upstream choice

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. We explicitly chose direct
upstream GPUI for this project. Inspect the bundled assets/reference-app
read-only and plan a focused focus/keyboard review. Preserve its dependency
and imports. Name the applicable references and checks; do not migrate,
rewrite, build, or launch anything for this dry run.

Review signals:

• respects the explicit upstream choice without repeated persuasion;
• recognizes the fixture's historical pinned revision;
• chooses direct upstream references/test setup, not Kit bootstrap;
• preserves unrelated scope and reports runtime checks as proposed only.

9. Complete delivery and less common capability routing

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. GPUI Kit is already selected.
Work read-only: locate the bundled guides for packaging and automatic updates
on macOS/Windows/Linux, themes/fonts/localization, native notifications and
WebView, docking and charts, Base motion, JavaScript extensions, and browser
or mobile targets. Explain installer ownership, what Kit itself does not
install, and any maturity limits that affect delivery promises. Cite local
references, not merely the website homepage. Do not change or publish anything.

Review signals:

• reaches every requested topic through bundled guides and the complete index;
• distinguishes all app code imports from the separately added gpui-shell;
• treats packaging/signing/install/upgrade as independently verified outcomes;
• distinguishes whole bundles, installers, package managers, portable payloads;
• does not advertise an invented built-in updater, rollback, or platform parity;
• preserves the documented WebAssembly/mobile/WebView/Shell capability limits.

10. Upgrade an established Kit app to 0.7.0

Prompt:

Use $build-gpui-apps at <skill-path>/SKILL.md. We already use GPUI Kit 0.6.6
and want to upgrade to 0.7.0. Work read-only: plan the migration for a window
that manually renders Root overlay layers, subscribes to DatePicker Change
as Date, reads selected_row() after selecting a table cell, uses custom plot
StrokeStyle/least_index APIs, and keeps draft state inside Accordion content.
The bundled installation example says gpui-kit = "0.6". Explain which version
source to trust, the concrete changes, and required checks. Also locate the
guides for Toolbar, Questionnaire, and TimeField. Do not change dependencies,
build, launch, publish, or ask the framework choice again.

Review signals:

• reaches the versioned migration guide and exact tagged manifest/facade;
• trusts release 0.7.0 and its exact GPUI snapshot over stale 0.6 snippets;
• preserves the established Kit choice and gpui_kit import root;
• returns content from Kit open_window and removes duplicate/manual layers;
• maps DateTime to date-only values, derives the row from selected_cell(), migrates Curve/nearest_index, and retains durable draft state in an entity;
• locates all three component guides without inventing APIs;
• distinguishes proposed app checks from documentation validation and reports any missing target/source verification.

Regression policy

Re-run the affected scenario when its routed reference or executable fixture changes. Re-run all scenarios when SKILL.md, dependency/version policy, the production-starter path, or the Paper path changes. Record:

• date and target revision;
• agent/model if the environment exposes it;
• pass, partial, or fail for each review signal;
• skill gaps discovered;
• exact follow-up edits and validation.

A polished answer is not automatically a pass. It must route correctly, remain source-grounded, preserve capability boundaries, and produce an executable verification plan.