Infographic generator

A pipeline for producing branded, operator-style infographics for social media. The design system is reverse-engineered from top product creators on LinkedIn — the look is clean grid structure + handwritten annotations + flanking emojis + signed footer. Outputs are ready for Twitter, LinkedIn, or Instagram.

The skill renders via OpenAI gpt-image-2 using the images.edit endpoint so the user's logos, screenshots, and avatar are honored as visual references rather than hallucinated.

What this skill produces

• Single-image infographics in 4:5 portrait, 1:1 square, or 16:9 landscape
• Consistent visual identity across posts via a saved theme color + avatar
• 12 layout templates (L1–L12) covering comparisons, before/after, stage flows, ranked lists, hero charts, framework grids, cheat sheets, ladders, myth/truth, process grids, visual metaphors, and annotated screenshots
• Reusable prompt patterns so each new piece takes ~5 minutes from idea to PNG

Project layout

When the skill is first used in a project, it creates a working directory:

infographics/
├── style.json          # active theme (filled at runtime: accent color, handle, avatar path)
├── assets/             # logos, screenshots, avatar (user-supplied)
├── outputs/            # rendered PNGs
└── prompts/            # one .py file per generated piece, kept for re-rendering

If this directory already exists, reuse the existing style.json (don't re-ask theme questions every session).

The 5-step workflow — always follow

When the user asks for a new infographic, run these steps in order. Don't shortcut.

Step 1 — Find the idea

If the user gave a clear topic, skip ahead. Otherwise:

• Surface 3–5 candidate ideas as a short table (idea title, why it'll resonate, suggested layout, suggested format).
• Pull from what you know about the user's work, recent context, or industry. Avoid generic ideas — specificity drives engagement.
• Recommend ONE pick with reasoning. Wait for confirmation.

Step 2 — Select layout + format

Pick one layout from references/style-guide.md (L1–L12). Justify briefly.

Then pick a Twitter/LinkedIn format:

Rotation rule: if the user is producing a series, alternate formats across posts to break feed monotony. Don't ship 5 portraits in a row.

Step 3 — Enrich with logos + images

Identify every brand, tool, or person mentioned in the idea. For each:

1. Brand favicons via Google's favicon service:

```

https://www.google.com/s2/favicons?domain=DOMAIN&sz=128

```

This is the most reliable logo source. Download with curl -sL -o assets/.png.

1. SVG-only logos: convert to PNG with cairosvg. Install inside a virtualenv to avoid touching the system Python:

```bash

python3 -m venv .venv && source .venv/bin/activate

pip install --quiet "openai>=1.0" "cairosvg>=2.7"

python3 -c "import cairosvg; cairosvg.svg2png(url='in.svg', write_to='out.png', output_width=512)"

```

If a venv isn't an option, ask the user before falling back to a system-wide install (e.g. pip install --break-system-packages) — it can affect their global Python environment.

1. User avatar: ask the user once for a photo path; copy into assets/avatar.png and reference in style.json.

1. Verify each PNG visually with the Read tool before proceeding. If a favicon came back as a generic globe (domain has no favicon), warn the user and either skip the logo or ask for an alternate URL.

Step 4 — Copywriting (ask the user)

The visual quality is bounded by the copy quality. Ask 3–5 targeted questions to fill the prompt. Tailor them to the chosen layout. Examples:

• Title (≤6 words, declarative or metaphorical) — what's the punch?
• Subtitle (handwritten line, one conversational sentence)
• Section/row labels specific to the layout (e.g. for L1: column headers + 5–7 row criteria)
• The takeaway (closing one-line rule that goes in the cream callout box)
• Tone: declarative / contrarian / playful / operator-honest

Do NOT generate copy yourself. Wait for the user. Their voice is the differentiator — the moment you write the copy, the post sounds AI-generated.

Step 5 — Generate

1. Compose the prompt: open references/prompt-templates.md, pick the right layout template, fill in all {{...}} placeholders with the user's confirmed copy, accent color, and reference image filenames.
2. Save the prompt as a .py file in infographics/prompts/.py so it's re-runnable.
3. Render via gpt-image-2: call scripts/generate.py (see that file for the exact API contract).
4. Read the output with the Read tool and show it to the user inline.
5. Note any rendering issues + offer one tightening pass (e.g. "Hostinger logo came out fuzzy — want me to swap to a cleaner source?").

Style invariants — never break these

These are what make the output look intentional rather than AI-slop. Skipping any one of them dilutes the whole brand.

1. Accent color carries the piece. Frame, accent words in the title, handwritten subtitle, hand-drawn arrows — all use the user's chosen accent. Read style.json for the active hex; if not yet set, ask the user.

1. Two emojis flank the title (top-left + top-right). One literal, one emotional. Use 3D Apple-style emojis, never flat ones.

1. Handwritten subtitle in a darker shade of the accent color (~25% darker), Caveat or Patrick Hand font. One sentence, conversational.

1. Footer signature pill: black rounded pill at bottom-center, small circular avatar (from assets/avatar.png) on the left, the user's handle on the right. Anchors the brand.

1. Closing panel is mandatory. Every piece ends with either a cream callout box ("rule of thumb"), a mint "final insight" panel, or a handwritten one-liner. Never end on raw data.

1. Use images.edit, not images.generate. This is what lets reference logos and the avatar be honored. generate will hallucinate them.

1. Don't say "purple" (or any color word) verbally in the prompt if it conflicts with the chosen accent hex. Models weight words higher than hex codes. Always describe the accent as the actual color word the user picked, AND include the hex.

Worked example — comparison infographic

User says: "Make a comparison infographic of Linear vs Jira."

Step 1 — Idea is clear. Skip to step 2.

Step 2 — Layout: L1 (comparison table) — head-to-head naturally maps to a table. Format: 4:5 portrait for density. Confirm with user.

Step 3 — Fetch logos:

curl -sL "https://www.google.com/s2/favicons?domain=linear.app&sz=128" -o assets/linear.png
curl -sL "https://www.google.com/s2/favicons?domain=atlassian.com&sz=128" -o assets/jira.png

Read both back to confirm they look right.

Step 4 — Ask:

• Title (≤6 words): your suggestion?
• Subtitle (handwritten, one line): the human take
• 5–7 comparison rows: which dimensions matter (speed, pricing, integrations, learning curve, mobile, ...)
• Takeaway: which side you actually pick + why
• Tone: declarative / contrarian / playful?

Step 5 — Compose the L1 prompt from references/prompt-templates.md, fill in placeholders, save to infographics/prompts/linear-vs-jira.py, render at 1024x1536, show output, note quirks.

Pitfalls to avoid

• Color word/hex conflict — if the user picks #006EFF (blue), don't leave the word "purple" or "violet" anywhere in the prompt. The model weights words higher than hex codes.
• Text density above ~12 cells — gpt-image-2 starts garbling. Shorten cell copy or split into two posts.
• Brand-name moderation — celebrity names, "Tesla", "Apple" (and other major brands sometimes) trip the safety system on images.edit. If you get a moderation_blocked error, swap the brand reference for a generic descriptor and retry.
• SVG logos — gpt-image-2 won't accept them. Always convert to PNG first.
• Outer frame missing — gpt-image-2 frequently drops the top edge of the outer accent frame. If a clean frame matters, composite it in Figma post-render (5 min job). Don't burn iterations chasing it.
• Tiny favicon source — Google's favicon API sometimes returns a 32px image. The model can still use it as a reference but the rendered logo may look fuzzy. Try &sz=256 first; if still small, source from the brand's press kit.
• Avatar drift — gpt-image-2 will sometimes alter the avatar photo. If brand consistency matters, composite the real avatar over the rendered output in Figma.

Security notes

This skill writes images, reads reference files, and uploads them to OpenAI. A few things to be deliberate about:

• API key scoping: OPENAI_API_KEY should be a project-scoped key with a spending limit. Image-2 calls are billable and can run up if the loop misfires.
• Reference images stay inside assets_dir: scripts/generate.py rejects absolute paths and .. segments in filenames. Don't try to work around the guard — if you need a logo from elsewhere, copy it into infographics/assets/ first.
• Output names must be bare basenames: out_name is validated for the same reason. Use names like linear-vs-jira, never paths.
• Prompt files: when invoking the CLI with python3 generate.py path/to/prompt.txt …, treat the prompt file as user-authored. Don't let an upstream agent point this at unrelated local files (the contents will be sent to OpenAI).
• Confidential assets: logos, screenshots, and avatars are uploaded to OpenAI under your account. Don't include private/internal product UI you wouldn't paste into ChatGPT.
• Persistent style state: infographics/style.json stores the user's accent color, handle, and avatar path. Treat it as branding metadata only — never store secrets there.

Reference files

When you need details, read these:

• references/style-guide.md — full visual rules: 12 layouts (L1–L12), color tokens, typography, recurring motifs, tone of writing
• references/prompt-templates.md — fill-in-the-blank prompt scaffolds, one per layout
• references/style.json — machine-readable design tokens, runtime-edited to record the user's theme

When the user has used this skill before

• Don't re-ask their theme color or handle — read infographics/style.json from their project.
• Reuse logos already in assets/ (don't re-fetch).
• Ask if they want to keep the same format as last post (rotation rule) or switch.

When to recommend Figma post-processing

Image-2 is great for the layout + composition but has known weak spots:

• Pixel-perfect outer frames
• Brand logos at small sizes (favicons under 64px)
• Specific typography (forcing a real font like Inter)
• Avatar fidelity

If any of these matter for a high-stakes post (launch announcement, sponsored content), generate the layout in image-2, then do a 5-minute touch-up in Figma to composite real logos + avatar + frame. Tell the user when that's worth doing.