Article Images

One skill, five modes, multiple brand profiles. Built for sustainability and research-led consulting work where image production is recurring and the cost of using AI image generation for structured visuals adds up. Replaces the clone-and-adapt-Python-script pattern from past articles, and replaces AI image generation for any text-on-background or chart visual.

Modes

How this fits with the other skills

• research-writing-assistant is the natural pair. When that skill writes a piece, it calls into this one in Stage 5 to generate the charts, featured image, and any stat cards the article references.
• gsc-audit outputs that warrant a chart (striking-distance keyword ladders, content decay timelines) can feed straight into chart mode here.
• wordpress-publish uploads the generated PNGs as featured images and inline media.

Brand profiles

Brand colours, fonts, attribution templates, and watermark behaviour live in brands/. One JSON per brand.

Recommended: use the setup wizard instead of editing JSON manually. Run python scripts/setup_brand.py (interactive) or pass --non-interactive with flags for CI use. The wizard validates colours, generates a default data palette, and saves a swatch preview. To verify a brand at any time, run python scripts/preview_brand.py --brand <slug> to render a multi-chart preview.

Brand auto-detection: when called from research-writing-assistant, infer the brand from the working directory or brief.json if present. Otherwise ask the user.

How to invoke each mode

Read the mode reference file when you reach that mode — not before.

For non-developer users, a unified CLI wraps everything:

python scripts/cli.py setup --slug acme --primary "#10b981"
python scripts/cli.py preview --brand acme
python scripts/cli.py featured --brand acme --title "..." --out featured
python scripts/cli.py stock --query "wind turbines" --count 3
python scripts/cli.py html --template title_card.html --brand acme --vars '{"title":"..."}' --output card.png

Working directory

All output goes to the calling article's working directory. When called from research-writing-assistant, that's {base}/[keyword-slug]-[date]/. When called standalone, ask the user where to save (default: current directory).

Environment

pip install -r scripts/requirements.txt once. Python 3.9+. Uses matplotlib, Pillow, requests, python-dotenv, html2image. The html-image mode requires Chrome / Chromium installed (which most desktops already have).

Source attribution rule

Every image with data or external content must cite its source. The brand profile defines the attribution template (font, colour, position). Charts get a footer line; stock photos get a markdown attribution block alongside the image file. Featured images that show only the brand's own content don't need attribution.

Why programmatic over AI image generation

For structured visuals — charts, stat cards, branded title cards, social images — AI image generation is the wrong tool:

• DALL-E 3 charges $0.04 to $0.12 per image. Gemini Imagen ranges from $0.02 to $0.24.
• Each image runs ~3 Wh of compute energy (per 2026 measurements).
• 3-5 iterations per image to land the right look.
• Output is generic, not brand-aware.

This skill produces equivalent or better visuals for ~$0 in cloud cost and ~0.01 Wh of local compute, with brand colours and fonts consistently applied. AI image generation stays useful for genuine photographic outputs (photorealistic compositions, custom photo edits) — keep it for that.

Non-commodity tie-in

Charts produced by this skill are inherently non-commodity content — they show original data analysis using the brand's own framing. When the research-writing-assistant outline includes a non-commodity-anchored section, that section is often a chart candidate.