art-direct

name	art-direct
description	Art direction for any content — reads text, PDF, Word, HTML, PPT, then proposes 2-3 creative directions with photography style, mood, and visual language. After selection, generates AI image prompts and visual briefs section-by-section. Use when the user shares content and needs visual direction, image sourcing, or creative direction for any material.

Art Direct

Turn content into visual direction. Point this at anything — a deck, a document, an essay, a brief, a webpage — and get back a creative direction you can actually execute.

When to Use

• User shares a file (any format) and needs visuals for it
• Developing visual identity for content before building/designing
• Translating written material into photography/illustration direction
• Creating image prompts for AI generation tools
• Art directing a presentation, document, report, or website
• Reviewing existing visuals against content intent (critique mode)

Supported Inputs

Read content from whatever the user provides:

Format	How to read
.txt, .md	Read tool directly
.html	Read tool, strip tags to extract text + structure
.pdf	Read tool with pages parameter
.docx	Extract via python3 -c "import docx; ..." or textutil -convert txt on macOS
.pptx	Extract via python3 -c "from pptx import Presentation; ..."
.rtf	textutil -convert txt on macOS
URL	WebFetch tool

If a format doesn't extract cleanly, ask the user to paste the text.

The Workflow

INGEST CONTENT → ANALYSE → PROPOSE 2-3 DIRECTIONS → USER SELECTS → VISUAL BRIEF + PROMPTS

---

Stage 1: Content Ingestion & Analysis

Read the full content. Extract:

1. Structure — What are the units? (slides, sections, chapters, paragraphs, pages)
2. Core themes — The 2-3 big ideas the content is actually about
3. Narrative arc — Does it build? Contrast? Layer? List?
4. Audience — Who receives this? What do they expect to see?
5. Tone — Authoritative? Inspirational? Intimate? Provocative? Technical?
6. Key moments — Which sections carry the most weight, demand the strongest visuals?
7. Existing visual language — If the content already has images, assess what's working and what isn't

Output a brief content summary before proceeding. Keep it tight — this is for alignment, not a book report.

---

Stage 2: Creative Direction Proposals

If house style template provided (--style <name>):

• Validate content fits the style
• Note any tensions and how to bridge them
• Skip to Stage 3 with adapted style guide

If no house style:

Propose 2-3 distinct visual directions. Each must be genuinely different — not three shades of the same idea. For each:

DIRECTION: [Name — a short handle like "Archival Authority" or "Warm Machinery"]

MOOD
What it feels like: [emotional quality in 2-3 words]
Energy: [calm / dynamic / tense / contemplative / electric]

PHOTOGRAPHY STYLE
Type: [documentary / editorial / conceptual / abstract / archival / illustrative]
Subjects: [what appears in the images]
Lighting: [quality of light]
Color treatment: [warm/cool shift, saturation, film stock reference]
Composition: [framing approach]

REFERENCE TOUCHSTONES
"Think [X] meets [Y]" — cite real publications, campaigns, photographers, or brands

WHAT THIS DIRECTION AVOIDS
[Specific clichés and visual tropes this direction rejects]

WHY THIS FITS THE CONTENT
[1-2 sentences connecting direction to content themes]

Present all directions. User picks one (or asks for a hybrid). Lock the choice.

---

Stage 3: Visual Style Guide

Once direction is selected, output the working style guide:

VISUAL STYLE GUIDE: [Content Title]
Direction: [Chosen direction]

PHOTOGRAPHY STYLE
─────────────────
Type: [Documentary / Editorial / Conceptual / Abstract / Archival]
Subjects: [What to feature — specific, not generic]
Composition: [Framing rules]
Lighting: [Light quality]
Color treatment: [Color approach, film stock if relevant]

MOOD & TONE
───────────
Primary emotion: [e.g., quiet confidence]
Supporting emotions: [e.g., warmth, precision]
Energy level: [Calm / Dynamic / Tense / Contemplative]

CONSISTENCY RULES
─────────────────
• [Shared quality all images must have]
• [Human subject guidelines]
• [Color palette anchors — hex codes]
• [Aspect ratio defaults]

CLICHÉ BLACKLIST
────────────────
• [Content-specific images to reject]
• [Generic tropes to avoid]
• [Overused metaphors for these themes]

AI GENERATION DEFAULTS
──────────────────────
Photography suffix: [standard prompt additions for photo-style generation]
Illustration suffix: [standard prompt additions for illustration-style generation]

---

Stage 4: Section-by-Section Execution

Work through the content in its natural units (slides, sections, chapters, key passages). For each:

Step 1: Interpret the section's job

What must the visual communicate? What's the emotional beat?

Step 2: Apply the Five-Lens Framework

Generate options through five lenses:

Lens	What it shows	When to use
Literal	The thing itself, shot with intention	Content is already specific
Human	People experiencing or doing it	Need emotional connection
Environmental	Setting, atmosphere, texture	Setting mood, transitions
Metaphorical	Concrete visual analogy	Making abstract tangible
Oblique	Abstract, unexpected angle	Provoking thought, standing out

Step 3: Output the Visual Brief

For the recommended lens (guided by the style guide's lens preferences), output:

SECTION: "[Section title or key line]"
VISUAL JOB: [What this image must do]
LENS: [Which lens and why]

CONCEPT
[2-3 sentence description of the exact image — specific enough
that a photographer could shoot it or a designer could find it]

AI GENERATION PROMPTS
─────────────────────
MIDJOURNEY:
[Full prompt with style suffixes, --ar, --v, --style flags]

DALL-E / GPT IMAGE:
[Natural language prompt optimized for DALL-E]

GEMINI:
[Prompt formatted for Gemini image generation]

IDEOGRAM:
[Prompt formatted for Ideogram, especially for any text-in-image needs]

SOURCING GUIDANCE
─────────────────
If searching (not generating):
  Search: [2-3 specific, refined search queries]
  Where: [Specific sources — see Source Guide below]
  Avoid: [What will come up that you should skip]

ALTERNATIVES
────────────
[1-2 other lens options briefly described, in case the primary doesn't land]

Step 4: For content with many sections

Don't generate all sections unprompted. Output:

1. The first 2-3 sections as examples
2. A summary table of all remaining sections with recommended lens and one-line concept
3. Ask which sections to develop fully

---

The Five-Lens Framework (Detail)

For any concept, five ways to see it:

Lens	"Digital transformation"	"Supply chain resilience"
Literal	Server room corridor, blinking LEDs	Cargo ship cutting through rough seas
Human	Developer's face lit by dual monitors at 2am	Dockworker's hands checking manifest in rain
Environmental	Empty office at dawn, single laptop glowing	Fog lifting off container yard at sunrise
Metaphorical	Old film projector casting light on blank wall	Spider web holding dew drops — tension + beauty
Oblique	Child's hand drawing a robot	Dominos frozen mid-fall, one glowing

The oblique lens is the hardest and the most valuable. It's the image that makes someone stop and think. Use it for hero images and opening sections.

---

Source Guide

Do not default to stock photo sites. Stock search produces generic results regardless of how specific your terms are. Instead:

Primary: AI Generation

The best match for precise creative vision. Generate exactly what the concept describes.

• Midjourney — Best for photographic realism and cinematic quality
• DALL-E / GPT Image — Best for conceptual and illustrative work
• Gemini — Good for diagrams, text-in-image, data visualization
• Ideogram — Best when image includes readable text or typography

Secondary: Editorial & Archival Sources

When you need real photography (historical, documentary, journalistic):

• Getty Editorial — Photojournalism, historical archives
• Magnum Photos — Documentary photography
• Library of Congress — US historical archives, public domain
• NASA Image Gallery — Space, earth science, technology
• Wikimedia Commons — Public domain, historical
• British Museum / Smithsonian — Historical objects and documents
• Internet Archive — Historical documents, publications, ephemera
• Google Arts & Culture — Museum collections, artworks

Tertiary: Curated Stock (when you must)

• Unsplash — Best for environmental/atmospheric shots, not people
• Pexels — Acceptable for textures, backgrounds, abstract
• Avoid for: People, business scenarios, technology in use, anything conceptual

For Specific Needs

Need	Best source
Historical technology	Smithsonian, Computer History Museum, Science Museum UK
Architecture	ArchDaily, Dezeen photography
Scientific	Nature journal imagery, NOAA, ESA/Hubble
Cultural	British Library, NYPL Digital Collections
Texture/material	Generate via AI — more control

---

Anti-Cliché Guide

Universal Blacklist

These images are invisible — viewers have seen them thousands of times:

• Handshakes (any kind)
• Lightbulb = idea
• Puzzle pieces connecting
• Person on mountain summit
• Hands holding globe
• Diverse team pointing at whiteboard
• Plant sprouting = growth
• Rocket = launch/speed
• Chess = strategy
• Maze = complexity
• Road diverging in forest = choice
• Iceberg = hidden depth
• Bridge = connection

The Reframing Technique

When you catch yourself reaching for a cliché:

1. Name the cliché — "I'm about to search for a lightbulb"
2. Ask: What does this concept feel like? — Not look like. Feel like.
3. Ask: What moment captures this for a real person? — Specificity kills cliché
4. Generate that instead

Example: "Innovation"

• Cliché: Lightbulb, circuit board, rocket
• Feels like: The moment before you know if it works
• Real moment: Engineer's hand hovering over a switch, not yet thrown
• That's the image

---

Critique Mode

When pointed at content that already has images (existing deck, webpage, document):

Step 1: Ingest & View Everything

Read all content. View every image. Do the work before speaking.

Step 2: Overall Visual Language Summary

Open with a top-level assessment of the visual language across the entire piece. Cover:

• What register are the images in? (archival, editorial, stock, mixed — name it)
• Is there a unified visual language? If not, how many competing registers are present?
• What's the gap between content intent and visual execution? The content is trying to say X; the images are saying Y.
• What's working and what isn't — broad strokes, not image-by-image yet

Keep this to a short, direct paragraph or two. This is the headline diagnosis.

Step 3: Section-by-Section Summary

For each section/slide/chapter, give a high-level summary — not an image-by-image table. For each section:

SECTION: [Title or key line]
CONTENT INTENT: [What this section is trying to communicate]
VISUAL EXECUTION: [What the images are actually doing — 1-2 sentences]
VERDICT: [Working / Partially working / Not working — and why in one line]
STRONGEST IMAGE: [Which one and why, if any]
WEAKEST IMAGE: [Which one and why — name the specific problem]

Only go image-by-image if the user asks to drill into a specific section.

Step 4: Recommendations

End with specific, opinionated recommendations — not open-ended observations. The format:

RECOMMENDED DIRECTION
─────────────────────
Register: [The specific visual register I recommend — e.g., "archival-documentary
          with warm color treatment" not just "pick a register"]
Why: [1-2 sentences connecting this to the content's actual themes and audience]
Reference: [Think X meets Y — cite real touchstones]

WHAT TO KEEP
────────────
• [Specific images that already work, and why they're the standard]

WHAT TO REMOVE IMMEDIATELY
──────────────────────────
• [Images that actively damage the piece — stock, wrong brand, wrong tone]

WHAT TO REPLACE
───────────────
• [Images that are weak/generic — with one-line replacement concepts]

CONSISTENCY RULE
────────────────
[The single unifying quality all images should share — stated as a rule
that can be applied as a yes/no test to any candidate image]

Be specific. Be opinionated. Don't say "commit to one register" — say "I recommend archival-documentary with warm tungsten color treatment, because this content is about heritage and the images need to feel like they were pulled from a real company archive. Think Bell Labs photography meets Kinfolk's material warmth."

Then ask: "Does this direction feel right? If so, I'll generate replacement briefs with AI prompts for every image that needs to change."

Step 5: Replacement Briefs (after user confirms)

Once the user agrees to the recommended direction, generate replacement visual briefs for every image flagged for removal or replacement. Use the full Stage 4 section-by-section format:

• Lock the recommended direction as the working style guide
• For each image to replace, output the full visual brief with:
  • Concept (specific enough to shoot or generate)
  • AI generation prompts (Midjourney, DALL-E, Gemini, Ideogram)
  • Sourcing guidance (where to find real alternatives if not generating)
  • One alternative lens option
• For sections that need additional images (currently too few), recommend how many and provide briefs

Output format: Generate all replacement briefs at once, numbered to match the original image positions. Export to a text file on the user's Desktop for easy reference and handoff.

Step 6: Handoff

After replacement briefs are generated, offer next steps:

• "Generate now" — Generate images via fal.ai (Flux 2 Pro) directly from the briefs
• "Export briefs" — Save all prompts and guidance to a file for use in Midjourney/DALL-E/external tools
• "Rebuild the deck" — Feed the style guide and replacement images into keynote-slides-skill to produce a revised version
• "Save as house style" — Lock the recommended direction as a reusable YAML template for future work with this brand

---

Image Generation via fal.ai

When the user selects "Generate now", generate images using Flux 2 Pro via the fal.ai API.

Requirements: $FAL_API_KEY environment variable must be set.

How to generate

For each image brief, run this via Bash:

curl -s "https://queue.fal.run/fal-ai/flux-pro/v1.1" \
  -H "Authorization: Key $FAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "<THE DALL-E/FLUX PROMPT FROM THE VISUAL BRIEF>",
    "image_size": "landscape_16_9",
    "num_images": 1,
    "safety_tolerance": "5"
  }'

This returns a JSON response with a request_id. Poll for the result:

curl -s "https://queue.fal.run/fal-ai/flux-pro/v1.1/requests/<REQUEST_ID>" \
  -H "Authorization: Key $FAL_API_KEY"

When status is "COMPLETED", the response contains images[0].url. Download it:

curl -sL "<IMAGE_URL>" -o "<OUTPUT_PATH>"

Generation workflow

1. Create an output directory: .art-direction/generated/ (in the project) or a Desktop folder
2. For each visual brief, take the DALL-E/GPT Image prompt (these work best with Flux)
3. Submit to fal.ai, poll for completion, download the result
4. Name files by section: section-01-heritage-grid.jpg, section-02-legacy-of-discovery.jpg, etc.
5. After all images are generated, display them for review using the Read tool
6. User can approve, request regeneration with adjusted prompts, or switch to a different lens

Image size options

image_size value	Use for
landscape_16_9	Presentation slides, hero images
landscape_4_3	Standard slides, documents
portrait_4_3	Vertical layouts, mobile
square	Social media, thumbnails
square_hd	High-res square

Batch generation

When generating multiple images, submit all requests first (don't wait for each one), then poll for results. This parallelizes the GPU work.

# Submit all requests, collect request IDs
for i in 1 2 3 4 5; do
  curl -s "https://queue.fal.run/fal-ai/flux-pro/v1.1" \
    -H "Authorization: Key $FAL_API_KEY" \
    -H "Content-Type: application/json" \
    -d "{\"prompt\": \"$PROMPT\", \"image_size\": \"landscape_16_9\"}" \
    | python3 -c "import sys,json; print(json.load(sys.stdin)['request_id'])"
done

# Then poll each request_id for results

Cost

Flux 2 Pro via fal.ai is pay-per-image. Typical cost is ~$0.05-0.10 per image. A full deck replacement (10-15 images) runs about $1-2.

Fallback

If $FAL_API_KEY is not set or the API is unavailable:

• Export briefs to file instead
• Note that the user can paste prompts into Midjourney, ChatGPT image gen, or higgsfield.ai manually

---

House Style Templates

Reusable visual directions stored in ~/.agents/skills/art-direct/styles/ as YAML.

name: "Style Name"
description: "One-line description with reference touchstones"

photography:
  style: documentary | editorial | conceptual | abstract | archival
  subjects:
    preferred: [list of subject types]
    avoid: [list of subject types to reject]
  lighting:
    preferred: [light quality description]
    avoid: [light quality to reject]
  composition: [framing rules]
  color:
    treatment: [color approach]
    palette_anchors: [hex codes]

mood:
  primary: [one emotional quality]
  supporting: [list of supporting ]
   [      ]


   [    ]
   [ ]
   


   [ ]
   [ ]


   
   


   
   

---

Quick Reference

Invocation	Purpose
art-direct	Full workflow — ingest content, propose directions, generate briefs
art-direct --style <name>	Apply existing house style template
art-direct --critique	Review existing visuals against content intent
art-direct --section "concept"	Quick single-section visual brief
art-direct --from-frontend <project>	Derive style from frontend-design project

Integration

Outputs can feed into:

• keynote-slides-skill — Visual briefs → slide imagery via Gemini generation
• branded-pptx-converter — Visual briefs → PowerPoint image slots
• frontend-design — Style guide → web design visual language