feat(outline): Design Brief mode — natural-language brief + structured media manifest

[wyuc]

Design Brief mode — natural-language brief + structured media manifest

Draft / RFC — opening for design discussion. Default behavior is unchanged; everything is behind an opt-in flag.

Motivation

The outliner today describes each slide with a terse {description, keyPoints}. Two problems show up downstream:

1. Weak layout signal. A bullet list under-specifies the page, so the slide-content model has to invent most of the layout. Richer, in-distribution guidance produces noticeably better and more consistent slides.
2. Decorative image spam. Because image intent only lives as prose ("add an icon", "a diagram here") and is not structured, the slide model emits an image element per mention. With image generation enabled those become per-card placeholders that mostly add noise; with it disabled they are dead boxes.

What this adds (opt-in via requirements.designBriefMode)

For every slide scene, the outliner now emits:

• brief — a detailed natural-language design brief (page goal, visual style + palette in words, region-based layout, the actual written-out content, focal emphasis). It becomes the authoritative slide-content input when present.
• media[] — a structured manifest of the images/videos the slide actually needs, unifying real assets and to-be-generated media:

  { id, source: 'asset' | 'generate', type: 'image' | 'video', prompt?, caption?, aspectRatio? }

  The brief references each item by id (gen_img_1, img_2), with mandatory two-way consistency (every brief id ⇔ a media entry).

The prompt also adds media discipline: most slides need zero images (icons/dividers are drawn with shapes; table/chart/latex carry information), and media is only requested when it carries irreplaceable visual information.

How it wires into the existing pipeline

• media[] generate items are bridged into the existing mediaGenerations (and asset items into suggestedImageIds) inside uniquifyMediaElementIds, so the current media-orchestrator dispatches generation unchanged. The id-uniquification now rewrites mediaGenerations, media[] and the inline brief references together, so all three stay consistent.
• slide-content consumes brief as the authoritative layout spec (new {{#if brief}} block); media flows through the existing generated-image instructions once bridged. Model-agnostic — no change to model routing.

Scope / files

8 files, ~150 lines: types (SlideMediaItem, SceneOutline.brief/media, UserRequirements.designBriefMode), the requirements-to-outlines prompt (gated blocks + two worked examples), the media[]→mediaGenerations bridge, and slide-content brief consumption.

Open questions for reviewers

• Flag surface: a request-level designBriefMode here — should it instead be a setting / per-stage option, and should there be a UI toggle (not included yet)?
• Whether to make the media[] manifest the canonical media field longer-term and derive mediaGenerations from it everywhere, vs the current additive bridge.
• Prompt-token budget of the gated blocks (only added when the mode is on).

Status

Types compile (tsc --noEmit clean). Not yet wired to a UI toggle (backend accepts the flag). Looking for feedback on the interface before fleshing out tests + UI.