Brand packages: structured branding guidelines linkable to tasks, projects, and asset/copy generation

[markhayden]

Problem

There's no structured concept of a "brand" anywhere in Bakin. Today the word only shows up as:

• Free-text in usageConstraints for the create-image runtime skill (plugins/images/defaults/runtime-skills/create-image/SKILL.md)
• A routing heuristic in plugins/images/lib/routing.ts that sniffs the literal string "brand" to pick a provider
• A UI placeholder ("brand-voice") in the workflows node-config drawer

So when an agent generates an asset, writes copy, or builds anything customer-facing, there's no machine-readable source of truth for voice, style, colors, logos, or reference material. Brand consistency depends entirely on whatever got pasted into a prompt.

Proposal

A first-class brand primitive — a named bundle of branding guidelines that can be linked to tasks, projects, workflows, etc., and consumed by generation paths.

Package shape (to be defined)

A brand bundle would include some mix of:

• Style guide — markdown describing voice, tone, do/don't rules
• Reference info — positioning, audience, terminology, boilerplate
• Visual identity — colors (structured palette), typography, logo files, reference images
• Arbitrary attachments — markdown docs, images, examples of on-brand output

This smells like a new agent-package kind (kind: "brand" alongside agent | skill-pack | workflow-pack | lesson-pack) — brands are content, not code, which is exactly the plugins-vs-packages split. Alternative: a dedicated brands plugin with its own storage under ~/.bakin/brands/. Tradeoff to work through in the spec — packages get install/lockfile/projection machinery for free; a plugin gets UI, routes, and search indexing for free. (Possibly both: plugin owns the UI/API, package kind is the distribution format.)

Linking

• brandId (or brands[]) on tasks, projects, and workflow instances — same pattern as other cross-object references
• UI affordance to pick a brand when creating/editing those objects

Consumption

• Dispatch: when a task has a linked brand, inject the brand context (style guide, palette, voice rules) into the agent's prompt/context
• Images plugin: linked brand's reference images + palette feed image generation (pairs with the reference-images issue, Images: support reference/context images in generate (and extra references in edit) #418)
• Copy/writing: brand voice doc available to any agent producing prose
• Hooks: expose brands.get / brands.resolveForTask via the HookRegistry so any plugin can consume brand context without direct imports

Open questions for the spec

• One brand per object or multiple (e.g., agency use-case)?
• Inheritance: does a task inherit its project's brand?
• How much is structured (colors as hex arrays) vs. freeform markdown?
• Versioning — do brand edits invalidate/flag previously generated assets?

Why

Ensures assets created, copy written, and anything agent-produced uses the right brand, voice, and visual identity — today that guarantee doesn't exist beyond prompt discipline.