perture19.8 KB
View saved version →
---
name: perture
description: Use a connected Perture brand for local frontend work, external host outputs, or explicitly requested Brand Page edits.
---
# Perture Design Agent
Use Perture as the source of brand intent and the local repository as the source
of implementation context. Preserve the user's product behavior, content,
permissions and chosen architecture.
## Select one mode
- **External output**: the user asks another host such as Sites, Canva or
ChatGPT to create a page, design or document using Perture Brand Page
references. Use the read-only reference workflow below. A page created in
Sites is an external output even though its result is frontend UI.
- **Build**: the user asks to create or change frontend UI in a local
repository. Read
[references/build.md](references/build.md).
- **Review**: the user asks whether existing UI follows the Perture brand. Read
[references/review.md](references/review.md). Repository access stays
read-only.
- **Fix**: the user asks to repair Perture findings. Read
[references/fix.md](references/fix.md).
- **Brand Editor**: the Pro owner asks to add or edit Brand Page colors,
typography, logos, fonts, illustrations, spacing or rounding. Read
[references/brand-editor.md](references/brand-editor.md). This mode does not
require a frontend repository or `prepare_ui_change`.
Build includes a scoped post-build review only when `automatic_reviews` is
enabled in the returned account behavior. Its safe repair loop also requires
`automatic_corrections`. Do not treat either automation as authorization for
unrelated cleanup. If the request is ambiguous between Review and Fix, begin
with Review and obtain authorization before any edit.
For Build, Review and Fix modes that run deterministic checks, read
[references/local-validation.md](references/local-validation.md) before running
the bundled validator.
## External output workflow
1. Use the connected Perture app to call `list_brands`, select the exact brand
with `set_active_brand`, and retain its `brand_lock_token`.
2. Before invoking Sites, another generator, or writing output source, call
`preflight_brand_output` with the exact output type. Include every capability
explicitly requested by the user in `required_capabilities` (for example
`illustration`, `photography`, `iconography` or `pattern`). A `blocked` result
ends generation: report its short `blocked_reasons` and ask the brand owner
to complete those Brand Page fields. Do not invent a missing style, use an
unrelated asset, choose a different brand, or try a generic host fallback.
The allowed hex list is exact; deriving lighter or darker shades is not
authorized. A `ready` result with `visual_authorized: false` permits only
an output without a new visual; `illustration_authorized: false` never
permits drawing a new illustration or substituting an unrelated asset.
Follow `visual_execution` when it is present. `official_asset` allows only
the mapped image unchanged. `generate_from_references` requires a new
original image guided by the returned reference images; those images are
inputs, not finished assets. If image generation cannot use the references
as visual inputs, stop and report the missing capability. A CSS product
mockup does not fulfill a required illustration.
This is a deterministic check before model-heavy generation,
not a second model review. If this tool is unavailable, stop and report an
incompatible Perture MCP version.
3. On `ready`, call `get_brand_reference_for_output` once with the actual output
type, logo background and available space. It includes the output guidance,
complete color list, logo choice and visual references. Use only the returned
official assets and brand guidance. Call `get_output_reference` only if a
specific field is absent from this package. For frontend work, call
`plan_brand_interface` once with every meaningful role:
controls, containers,
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 13 — file references are metadata, not package contents.