# Prompt examples

Use these prompts with any [connected MCP agent](./connect-ai-agents.md).
Replace bracketed values with captured product, version and page identifiers.
Select platform type and environment IDs from `list_platforms.options.platforms[]`,
which includes browser and native environments. Use the returned IDs rather than
assuming desktop/mobile is an exhaustive list. Use only captured pairs from
`list_platforms.platforms[]` for reconstruction; supported options do not imply
stored data.
## Find a product and version

> Use Reconstruct MCP to list products. For `[product]`, list captured platform/environment pairs, then versions for platform `desktop` and environment `web`. Select one version and summarize its pages and artifacts. Do not combine versions or environments.

Expected path: `list_products` → `list_platforms` →
`list_versions(platform="desktop", environment="web")` →
`get_platform_summary(platform="desktop", environment="web")` → `list_pages`.

`list_pages.views[]` gives exact stored profile keys. Select one of those keys for
`view`; it may be `desktop`, `mobile`, a numbered profile or a device label.

## Rebuild a page from a capture

> Inspect `[product]`, version `[version]`, environment `[environment]`, page `[page]`. Discover the stored profiles with `list_pages`. For `[view]`, call `get_reconstruction_bundle` once per requested vertical fragment of at most two viewport heights. Consume the JSON with code outside model context and copy the HTML starter, JSON and referenced media into one directory, preserving filenames. Use the structured tree, text, geometry and styles as evidence. Verify `starterLimitations`, `captureTruncated` and `selectionComplete`; report missing data and inferred behavior.

The default response contains file references, not the complete IR. For HTTP,
server filesystem paths are remote: download the returned URIs using the MCP
server origin and Bearer token.

## Recreate one component

> For `[product]`, version `[version]`, environment `[environment]`, page `[page]`, stored view `[view]`, find `[component]` using `get_page_outline`. Read one `get_section_bundle` for its parent region, or batch selected regions with `get_page_bundles`. Use preview text first. Request full text or individual component layers only when the bundle is insufficient. Resolve media with `get_element_asset` and fonts through `get_fonts` plus `get_capture_asset`. Implement the component and identify which behavior is inferred.

Component boundaries are heuristic. Nested outline leaves are structure hints;
avoid requesting one bundle per leaf. Image tools provide optional final visual
checks, while structured IR provides content and layout evidence.

## Audit navigation and flows

> For `[product]`, version `[version]`, environment `[environment]`, stored page/view `[page]`/`[view]`, identify captured destinations with `get_page_transitions`. Inspect `list_scenarios`, scenario manifests and recorded steps in that same environment. Separate discovered links from actions that were actually recorded. Check replays when available.

## Compare desktop and mobile

> For `[product]`, version `[version]`, environment `web`, list the stored profiles for `[page]`. Compare one desktop profile and one mobile profile using summaries, IR and measured geometry. Identify changed layout, moved or missing content, and capture limitations. Use images only for optional visual verification; do not infer a missing mobile capture from desktop data.

For native comparisons, choose explicit captured pairs, for example
`desktop/macos` and `mobile/ios`, and keep each environment's evidence separate.

## Inventory design assets

> For `[product]`, version `[version]`, platform `[desktop-or-mobile]`, environment `[environment]`, inspect `get_technology_stack`, `get_fonts`, `get_themes`, `list_product_assets` and `list_archived_assets`. Prefer captured original files through `get_capture_asset`; use origin URLs only when stored bytes are unavailable. Identify fonts, themes, media and detected technologies. Disclose substituted resources and uncertain source-code details.

Capture currently keeps up to **2,500 visible elements**. Hidden UI, hover/focus
states and dynamic behavior may be absent or partial. The default crawl uses the
seed locale; `captureAllLocales=true` can include translated routes. Report actual
`truncated` and capture-limit metadata rather than assuming full coverage. See the
[tool catalog](./mcp-tools.md) for parameters, access requirements and usage limits.

## Submit and monitor a product request

> Get the available request options and submit an add request for https://example.com/
> using desktop/web if that pair is available. Keep the returned request ID, check
> its status, and wait for changes using wait_product_request. Continue until it is
> approved or rejected; report the request decision. Approval does not indicate
> that the new capture version is published.

> Submit an update request for product `<product-key>` using the selected
> request platform/environment. Use the existing product source URL. Keep its
> request ID and monitor it; do not create another request after a wait timeout.