pf-asset-orchestrator

name	pf-asset-orchestrator
description	Reads asset prompts from 10_Content.md, routes each to the right generator (Higgsfield for images/video, Napkin.ai for infographics, Flaticon for icons), saves outputs to WIP/[deck]/assets/, and updates 10_Content.md with the resolved local paths. Polls long-running jobs and reports back.
type	skill
plugin	presentation-factory
tags	["type/skill","plugin/presentation-factory","scope/specialty","topic/asset-generation"]
status	active

When to invoke

Called from the Compose stage orchestrator before the slide composer runs. Assets need to exist before slides are composed because the composer needs resolvable image paths.

Required context

Toolchain: ../../references/presentation-factory-toolchain.md. Foundations: ../../references/presentation-factory-foundations.md.

What this skill does

Stage 1: Parse asset prompts from content file

Scan WIP/[deck]/10_Content.md for asset blocks. Three types:

• Image prompt blocks (Higgsfield) → output PNG/MP4 to assets/images/
• Infographic blocks (Napkin) → output PNG/SVG to assets/infographics/
• Icon blocks (Flaticon search) → output SVG/PNG to assets/icons/

Each prompt is associated with a slide number (the section header above it).

Stage 2: Image generation (Higgsfield)

For each image prompt, call the Higgsfield MCP generate_image tool with the prompt text. Default settings:

• Model: nano banana 2 for stills (general-purpose, fast)
• Alternative: chatgpt image 2 for more photorealistic
• Aspect ratio: 16:9 for full-bleed, 4:3 for two-col layouts, 3:4 for portraits
• Resolution: 1920×1080 for hero, 1280×720 for inline

For video prompts (rare; only for openers or specific story slides), use seedance 2.

Workflow:

1. Submit prompt → returns job ID
2. Poll job_display until ready (or use the listener pattern if the MCP supports it)
3. Download the result via show_medias or direct URL
4. Save to WIP/[deck]/assets/images/slide-NN-image.png (or .mp4 for video)
5. Update 10_Content.md: replace the **Image prompt** block's path with **Image**: ../assets/images/slide-NN-image.png

If generation fails (rate limit, content policy, etc.):

• Log the failure with the slide number and reason
• Don't block the rest of the asset generation — continue with others
• Surface failures to user at the end with options: retry with adjusted prompt, manually source, or use a placeholder

Stage 3: Infographic generation (Napkin)

Napkin.ai does not have an MCP. Two workflows:

Path A — Manual (default):

1. For each infographic block, write the Napkin-ready description to WIP/[deck]/assets/infographics/slide-NN-prompt.md
2. Tell the user: "Open Napkin.ai, paste these prompts, save results to the infographics folder."
3. After user signals done, verify each expected file exists

Path B — Browser MCP (if mcp__Claude_in_Chrome available and user wants to skip manual):

1. Use the browser MCP to drive Napkin.ai
2. Submit each prompt, screenshot result, save to disk
3. This path is slower and more error-prone — only use if user explicitly opts in

Stage 4: Icon sourcing (Flaticon)

For each icon block:

1. The block contains a Flaticon search query and a style preference
2. Either: (a) drive a browser search via mcp__Claude_in_Chrome and download the chosen icon, or (b) write the search query to WIP/[deck]/assets/icons/slide-NN-needed.md and ask user to source manually
3. Save SVGs to WIP/[deck]/assets/icons/
4. Update 10_Content.md with the file paths

Pick ONE icon family for the whole deck. Don't mix Lucide / Phosphor / Flaticon styles within a deck — design:design-iconography covers this.

Stage 5: Validation

After all generation runs:

• Count expected assets (parsed from content) vs. actual assets on disk
• Surface any missing assets to the user
• For images that came out wrong (visible to LLM via small thumbnail check): offer re-generation

Stage 6: Update content file

Rewrite 10_Content.md so every asset reference points at a local path. The composer skill needs concrete file paths to substitute into the templates.

Higgsfield prompt patterns that work

From the design:design-imagery skill — these are the principles to follow:

• Specific scene, not generic: "wide shot of a healthcare nurse at a workstation, mid-conversation with a colleague, natural light, slight motion blur on background" — NOT "professional healthcare worker"
• Mood line: always include a mood ("collaborative," "focused," "warm") — single word per direction
• Style line: documentary / editorial / abstract / illustrative — pick one and hold
• Light direction: "natural daylight from left," "warm low-angle morning light" — light is the easiest brand differentiator
• Avoid clichés: explicitly negate ("no handshakes, no pointing-at-screens, no fake diversity tableaux")
• Color palette: "warm-cool contrast" or "muted earth tones with one accent" — guide the color story
• Resolution: always specify (1920×1080 for hero, 1280×720 for inline)

Anti-patterns

• Don't generate generic stock-photo-style imagery (suited men pointing, fake diversity tableaux). Every image gets a specific scene description.
• Don't mix icon families across a deck.
• Don't silently swap a failed generation with a placeholder without telling the user.
• Don't generate assets if the user hasn't signed off on 10_Content.md — image generation has cost. Confirm content first.
• Don't request more than ~12 unique images per deck — that's a strong signal the deck is over-decorated. Push back via user prompt.