skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
yschimke/skills52 installs

compose-preview

Render a Jetpack Compose or Compose Multiplatform @Preview to PNG and look at it. Use when asked to render, show or check what a composable, @Preview or Material/Wear library component looks like, to verify a Compose UI edit, to compare before/after, or to review previews against design guidelines. Also to run an Android APK without an emulator.

How do I install this agent skill?

npx skills add https://github.com/yschimke/skills --skill compose-preview
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill 'compose-preview' provides tools for rendering Jetpack Compose previews. It includes installation scripts that fetch components from the author's GitHub repository and other technology-specific services. The skill uses a rendering daemon to process Kotlin code into images. No malicious patterns were detected.

  • Socketpass

    No alerts

  • Snykpass

    Risk: LOW · No issues

What does this agent skill do?

Compose Preview

Render @Preview composables to PNG without Android Studio (Android via Robolectric, CMP Desktop via Skia). CLI, Gradle plugin and renderers ship from compose-ai-tools; the MCP server from compose-preview-server.

Render first

For a request to render or show a preview, follow the short loop below. For a requested guidelines review, use the review checklist to choose the local, hosted or UI Builder lane before invoking tools. Published catalogs and server-homed designs use hosted tools first; a missing local checkout does not require cloning, Gradle or a cloud environment. If tools are unavailable, report blocked evidence and follow the integration's setup guidance.

With the local MCP server attached (compose-preview-server 3.78.0+), its initialize instructions say how to call it: render_preview preview=<FunctionName>, no exploring or registering first, no hand-built mocks. Follow them. This skill adds the budget: a routine render is at most three tool calls, with no base64 in the reply and images only when you need to see them.

Your first call is the render. Don't grep for the preview or read its source first: the server finds it by function name. Look things up only if the render fails.

  1. Render. If your client reads local files (Claude Code, Codex, Gemini CLI, OpenCode, Antigravity), pass inline=false. The result is pngPath, sha256, dimensions and changed, with no image tokens. If it lists otherMatches, render one only if the person asked about it. If it carries a variantChoice whose message says to ask (several previews matched and the host showed no chooser), the person picks, not you: reply with the rendered match and every entry in choices as a numbered list, ask which one they meant, and stop. Don't pick one or shorten the list to "other variants are available". A grid result or a declined choice needs no question.
  2. Look only when you need to. Read pngPath with your file reader before you describe or judge the UI; that is what the person sees. Skip the read when changed or sha256 already answers the question, such as "did my edit land?". Describe only what you saw. If you can't view images here, say so plainly. Never claim a visual check you didn't make. changed only means the pixels differ from an earlier render; it can be true on the first render of a session and it says nothing about how they changed. To say an edit looks right ("the title is bigger", "verified", "fixed"), read pngPath first. If you didn't, or a render failed, say the result is unverified.
  3. Reply briefly, with pngPath so the person can open the same image. Don't write "the image above": in a terminal or print-mode harness the person doesn't see images your tools return. Point at pngPath instead.

Library components come from the catalog. When the person names a library component (Material 3, Wear M3 EdgeButton, …) rather than one of their own previews, use the hosted catalog server (compose-preview-catalog): catalog_list_previews, then read the published snapshot with resources/read. Request live access only to change knobs with catalog_render_preview. Never add preview files to the person's project to show a library component unless they ask for one. See catalog-mcp.md.

Sweeps use hashes. For more than one render (variants, devices, font scales, locales, a before/after check), pass observe=hash or use render_matrix with contactSheet: false, which returns per-cell hashes only. Fetch pixels only for the final screen or the cells whose hash moved, or that you must judge by eye: one render_preview per such cell (its overrides or uri, inline=false) and one read of its pngPath.

Design guidelines review. When asked to review a preview or catalog design against Android design guidance, read design-guidelines.md. It covers keyless agent review, published results and the CLI's compose-preview review loop, including explicit surface selection, required pictures and incomplete coverage. A project with no guidelines of its own reviews against the shared rule packs chosen by what each preview draws (widgets need the widget packs and a @LauncherWidgetPreview), not a component catalog. It is in the default skill bundle; the optional PR-review bundle is not required.

Hand multi-render reviews to design-reviewer. If a design-reviewer subagent is available (the compose-agent-plugins plugins ship one), delegate accessibility, font-scale, round-device and other matrix checks to it so the images stay out of your context, and relay its verdict and paths. Give it the subject, home/revision, catalog rules and requested coverage. Without one, run the same checklist yourself and report that delegation was unavailable. For a routine hash sweep, keep contactSheet: false. A requested visual guidelines comparison needs pixels: inspect its required frames even in your own context when no reviewer exists. Prefer only the relevant individual images or a bounded comparison sheet, and never claim visual coverage from hashes alone.

After a source edit, call render_preview again. notify_file_changed is optional. If the result says it is stale, make exactly one more call with force: {"reason": "<why>"}. Never run ./gradlew, clean, or delete build/ dirs to chase a render; if a forced render is still wrong, read mcp.md § Troubleshooting (start with compose-preview mcp doctor).

Details on request. Add details: ["a11y", "layout"] only when the person asks about accessibility or layout. observe=png returns inline pixels for clients that can't read files.

No MCP tool? Use the CLI: compose-preview show --json --filter <Name>, then read the entry's pngPath. Stale: --force=<reason>, once. See cli.md. On either path, only output from the render tools is a render: never a hand-built HTML, CSS or SVG mock. Report a failed render with its error.

Running the whole app, not one composable? To drive or test an APK the way an emulator would (navigation, deep links, runtime permissions, foldable postures, a second display, offline or slow networks, BLE peripherals, home-screen widgets, a Wear app with its tiles, ambient mode and phone sync), or to see an edit running with live reload (launch --project, reload), use compose-preview launch <app.apk> or the MCP simulator_* tools: Robolectric's simulator, no emulator or device. Read simulator.md first; a @Preview render stays the cheaper check for a single screen.

Not previewable (takes a ViewModel or injected service)? Propose extracting a stateless inner composable and preview that (state-hoisting.md).

Install / update (only if missing)

Check with compose-preview --version. If it's missing, run this skill's stub, bash <this skill's dir>/scripts/compose-preview --version: its first run installs the CLI and puts it on PATH. Every other route, updating, and the Gradle plugin are in setup.md; MCP wiring (the local server and the one hosted URL) is in mcp.md.

Working with platform skills

An official platform skill, such as Android's Wear Compose Material 3 skill, owns API, dependency and migration choices. This skill only renders and verifies:

  • Defer. Follow the platform skill for which component, API or version to use. Don't treat this skill's references as a second API guide.
  • Changes are expected. During a migration a changed hash or image is information, not a regression. Only a render failure or a new accessibility error blocks. Don't restore old screenshots or tune the UI to match them.
  • Mind the catalog version. A hosted catalog renders one fixed library version, which may differ from the project's libs.versions.toml. See catalog-mcp.md § Choosing a component.

Migration workflow. Before editing, find_previews_for_file for each file and record a baseline with render_preview observe=hash. Make the platform skill's edits, re-render, and report failures and new accessibility errors (get_preview_data kind=a11y/atf) as blockers. Show a before/after for one to three key screens, and check a small round device and the largest font scale on Wear. Details: wear-ui.md § Verification workflow.

Reference index

Read only what the task needs.

TopicRead
Install, update, which skills, apply the Gradle plugin, zero-code init scriptsetup.md
CLI commands, --filter/--id/--force, counts buckets, render filenames, SVG output, Gradle tasks, build-brief for non-preview buildscli.md
Which MCP server (local vs hosted https://preview.coo.ee/mcp); local server setup, tools, history, multi-workspace, troubleshootingmcp.md
Interaction loop: semantic refs, observe modes, diff_semantics, crop, render_matrix (variants/matrix), record → test, failure kindsagent-loop.md
Accessibility (ATF, compose-preview a11y)a11y.md
Per-render data (a11y, layout, recomposition, SVG)data-products.md
Multi-preview, GIFs, @SettledPreview, scrolling capturescapture-modes.md
Run an APK in the simulator (launch, simulator_*): drive an app, adb equivalents, form factors, folds, multiple displays, home screen and widgets, permission ask policy, Play services statesimulator.md
Simulator network modes, real-traffic gating, Wi-Fi requests, Bluetooth / BLE / GATTsimulator-network.md
Simulator on Wear: watch face, tiles, notifications, ambient and screen power, --pair, --data-layer agent, ChannelClient, companion devicessimulator-wear.md
Simulator edit-run loop: restart, update, live reload (--project, reload, --live), --adbsimulator-iterate.md
Wear Compose UI / Wear Tileswear-ui.md, wear-tiles.md
CMP :shared modulescmp-shared.md
Android XML drawables / iconsresource-previews.md
Runtime permissions per renderruntime-permissions.md
Editable override knobsoverride-knobs.md
Display filtersdisplay-filters.md
Remote Compose, compose-preview rcremote-compose.md
Hosted catalog MCP tools and grant flow (serve, no checkout)catalog-mcp.md
Gated server access (compose-preview auth request)server-access.md
Agent allowlists, staging PNGspermissions.md
Cloud sandboxesagent-cloud.md
VS Code extension (humans)vscode.md

CI and PR review are the opt-in compose-preview-ci and compose-preview-review skills; setup.md § Which skills adds them.

Offer the next step

When one of these signals is visible, you may end a reply with its one-sentence offer. At most one offer per reply, and never repeat an offer the person ignored.

SignalOffer
A PR that changes UI, and the repo has no preview CI"I can set up CI that posts before/after renders on PRs." Needs the review skills: scripts/install.sh --skills compose-preview-review,compose-preview-ci, or /plugin install compose-review@compose-agent-plugins.
Reviewing a UI PR"I can render base and head and diff them for this review." Same review skills as above.
They want to browse every preview themselves"compose-preview browse or the VS Code extension shows them all."
They need navigation or a running app"I can run the whole app without an emulator." See simulator.md.
They want to design a new screen rather than edit code"The UI Builder can design it with you first: https://ui.coo.ee/start."

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/yschimke/skills/compose-preview">View compose-preview on skillZs</a>