Download the design authoring skill

Bee Adapt Design helps a design author keep an interactive mock’s source, identify its screens and states, and hand it to an implementation agent with a compact manifest. Download the portable skill, enable it in your authoring host, and select it when asking for a design.

Download

Download Bee Adapt Design 1.0.1 (ZIP, 16,089 bytes) · SHA-256 checksum file

6d82cf356e8ce6ec3a07219bb95ab003f828d79b6f101b8d864466c9edf7156b

This is an independently versioned authoring preview. It supplies instructions and examples; it does not install the Bee CLI, a plugin, an MCP server, an exporter or a visual comparison tool. Bee’s own CLI/plugin design tooling shipped in 1.0.0 — bee design, bee design mcp, bee visual, and the Bee plugin’s bundled adapt skill; this download remains the separate, CLI-independent authoring preview for a design host that does not install the Bee CLI or plugin.

The public ZIP is also the source package: it contains one bee-adapt-design/ folder with six files: SKILL.md, standalone design instructions, two references and two synthetic examples. Read the files before enabling the skill. To check the downloaded ZIP and checksum file together on macOS or Linux, run this in their download directory:

shasum -a 256 -c bee-adapt-design-1.0.1.zip.sha256

Then inspect the files locally, for example with unzip -l bee-adapt-design-1.0.1.zip and unzip -p bee-adapt-design-1.0.1.zip bee-adapt-design/SKILL.md.

Expect bee-adapt-design-1.0.1.zip: OK.

Install and select in Claude Design

The following account-upload and Design-selection path was observed on 26 September 2026. Labels can vary; Claude’s skill guide describes current account prerequisites and upload controls.

1. Download the ZIP above. Keep it zipped for Claude’s upload.
2. In Claude, open Customize → Skills → Yours → Add → Upload skill and choose the ZIP. Some interfaces show + → Create skill → Upload a skill.
3. Confirm bee-adapt-design appears in your account’s skill list and is enabled.
4. Open Design. In its prompt, choose + → Skills → Your skills → bee-adapt-design. Confirm it is selected before submitting the request.
5. Describe the design, required interactions and states. Ask for the working mock and its source handoff together.

For example, with the skill selected:

Create an interactive task-list mock using the supplied brand and design system.
Keep one review canvas. Include populated, empty, loading and error states,
plus an add-task dialog. Apply bee-adapt-design and return the actual export,
stable screen/state/component/instance markers and a compact handoff manifest.
List unavailable dependencies and checks you could not run.

If upload controls are absent, check the account/organization’s skill permissions and code-execution setting using Claude’s guide. If the skill is enabled but absent from Design’s selector, do not assume it is active. You can instead attach or paste the ZIP’s CLAUDE-DESIGN-INSTRUCTIONS.md as explicit project instructions; that is a manual instruction path, not proof of account-skill activation.

Verification status: version 1.0.0 was uploaded, selected in standalone Design and used to generate a mock. That run could not read the supporting files. Version 1.0.1 puts the essentials directly in SKILL.md so references are optional. Its package checks passed; an updated Claude Design run with 1.0.1 is not yet verified.

Install in a Codex project

Keep your existing Codex configuration, model choice and authentication. This installation adds only the portable skill to your project.

1. Download and verify the ZIP, then extract it. Check that it contains one bee-adapt-design/ folder with SKILL.md and the five supporting files.
2. Copy that complete folder into your project’s .agents/skills/ directory. The entrypoint must be .agents/skills/bee-adapt-design/SKILL.md, with the reference and example folders beside it. Avoid an extra ZIP-name folder between skills/ and bee-adapt-design/.
3. Start a new Codex session in that project. Open the skill picker and confirm that bee-adapt-design is available before claiming it is installed.
4. Select the skill or explicitly invoke $bee-adapt-design with the design or adaptation request. Ask Codex to read the local entrypoint and the supporting references relevant to the task.

For example:

$bee-adapt-design Create an interactive task-list mock using this project's
design instructions. Preserve one review canvas and provide its actual source,
stable markers and compact handoff manifest. Check the required interactions
with the available local tools and report exactly what ran and what remains unverified.

If the skill is absent, check the folder layout and open a fresh session. Passing the SKILL.md path as ordinary task instructions is a useful fallback, but does not prove native skill discovery. A locally generated preview is distinct from a Claude Design canvas.

What to request in the export

• Actual source and assets: preserve the working review canvas, HTML/CSS/SVG, fonts, scripts and declared dependencies. A screenshot alone is insufficient for source reuse.
• Stable markers: data-bee-screen, data-bee-state, data-bee-component and unique data-bee-instance values on existing semantic roots, without extra layout wrappers.
• Compact handoff: design revision, entry file, dependencies, selectors, required states, deterministic fixtures and setup → action → visible outcome. Use a sidecar manifest or inert bee-design-handoff JSON in the canvas when supported.
• Explicit gaps: distinguish design approval, implementation readiness and real API behavior. Report actual checks and missing coverage; let the receiving tools calculate hashes from exported bytes.

Check that the export opens and its required interactions work in the intended environment. Supporting files may be unavailable in a cloud authoring host; the skill’s entrypoint is self-contained, and the author should report what it could actually read. The portable manifest is a handoff convention, not a promise of compatibility with an installed Bee command. Framework/API integration and local visual checks remain separate work. No automatic pixel identity or measured token-saving percentage is claimed.