# Reconstruct skills

Reconstruct skills give an AI agent a focused workflow for a product task. Each skill contains instructions and, where needed, supporting templates. Choose the skills relevant to your work; using one does not automatically run the entire product lifecycle.

You can install skills directly from the Reconstruct website, download them manually, or read them through a connected MCP client. A marketplace account is not required.

## Available workflows

| Skill | Use it to |
|---|---|
| `reconstruct-inspect` | Gather captured interface evidence, identify states and flows, and agree on scope. |
| `reconstruct-design` | Define design tokens, components, layouts and interaction states. |
| `reconstruct-build` | Implement the agreed interface and behavior in your project's stack. |
| `reconstruct-verify` | Check behavior, accessibility and visual fidelity against the available evidence. |
| `reconstruct-backend` | Plan and implement server behavior, contracts and integrations. |
| `reconstruct-entrepreneur` | Research customer problems and validate product opportunities. |
| `reconstruct-brand` | Develop a consistent product identity and brand assets. |
| `reconstruct-launch` | Prepare launch materials, listings and supporting evidence. |
| `reconstruct-deploy` | Prepare, deploy and verify an authorized release. |

Captured interfaces provide evidence about visible structure and behavior. They do not expose the original application's private backend implementation. The backend workflow uses your requirements and available contracts to design that behavior.

## Install from the website

Use the `skills` CLI with the website's root address:

```sh
npx skills add https://reconstruct.dev
```

Preview the available skills without installing:

```sh
npx skills add https://reconstruct.dev --list
```

Install a selected workflow into Codex:

```sh
npx skills add https://reconstruct.dev --skill reconstruct-inspect -a codex
```

Install all nine into Claude Code:

```sh
npx skills add https://reconstruct.dev --skill '*' -a claude-code
```

The CLI reads our public discovery catalog and downloads an archive for each selected skill. Each archive includes its instruction, supporting files, LICENSE and NOTICE. `npx` requires Node.js and npm; connecting to the remote MCP server does not require Node.js or a .NET SDK.

After installation, ask your agent to use the selected skill for your task. Skill selection and activation depend on your client. Skills that need captured data also require an [authenticated Reconstruct MCP connection](./connect-ai-agents.md).

### Manual installation

1. Download [the complete skill bundle](https://reconstruct.dev/downloads/reconstruct-skills.zip).
2. Extract the archive and choose the desired folders under `skills/`.
3. Copy those folders into your agent's supported skills directory. Keep the bundle's LICENSE and NOTICE with the installed distribution.
4. Reload your agent's skills if required by the client.

The complete bundle contains the nine workflows and their templates. For the CLI, use the website root address shown above; the complete ZIP is the manual download option.

## Read skills through MCP

Connect your agent to **`https://reconstruct.dev/mcp`** using Streamable HTTP and OAuth or a personal MCP token. See [Connect an AI agent](./connect-ai-agents.md) for client-specific setup.

Clients that support MCP resources can discover and read the same instructions through the existing connection:

1. Call `resources/list` to discover the catalog and skill files.
2. Call `resources/read` with the catalog URI:

   ```json
   {"uri":"skill://reconstruct/catalog.json"}
   ```

3. Choose a workflow from the catalog and read its advertised entrypoint. For example:

   ```json
   {"uri":"skill://reconstruct/reconstruct-inspect/SKILL.md"}
   ```

4. Read supporting files from the resource URIs advertised by `resources/list`. The inspection template, for example, is:

   ```json
   {"uri":"skill://reconstruct/reconstruct-inspect/assets/scope.md"}
   ```

Connecting MCP makes the resources available to the client. It does not automatically install skills into the local filesystem. If your client exposes tools but does not support reading MCP resources, install the skills from the website instead and keep the MCP connection for captured data.

## Agent discovery

The public [skill discovery catalog](https://reconstruct.dev/.well-known/agent-skills/index.json) lists each skill's name, description, archive URL and SHA-256 digest. Archives include supporting templates so installation preserves relative file references.

Individual instructions and supporting files are also available below `https://reconstruct.dev/.well-known/agent-skills/`. For example:

- Inspection instructions: `https://reconstruct.dev/.well-known/agent-skills/reconstruct-inspect/SKILL.md`
- Inspection scope template: `https://reconstruct.dev/.well-known/agent-skills/reconstruct-inspect/assets/scope.md`

Discovery is public and does not grant access to captured product data. Clients decide which instructions to fetch, activate or install. Marketplace listings can provide another way to find the package; direct website and MCP access remain available independently.

## Context and usage

Start with the catalog and load only the workflow and supporting files needed for the current task. Reading every skill adds instructions to the agent's context and can increase model token usage. Installing all nine does not require activating all nine at once.

Public discovery, downloads and MCP skill resource reads do not consume Reconstruct capture context reads. Capture tools called while following a skill use the normal [subscription limits and usage rules](./mcp-tools.md#usage-limits). Model token usage is billed or limited by your AI provider.

## Updates and license

Repeat installation from the website to obtain the currently published package. Website archives and MCP resources are released with their respective services; a locally installed copy does not change automatically when the server is updated.

Reconstruct skills are distributed under the MIT license. LICENSE and NOTICE include attribution for material adapted from Replica and for Reconstruct contributions. Review those files when redistributing the package.