Power Upgrade: Claude Code + Image Generators

The Problem

You run a company by yourself, or close to it, and the deliverables never stop. A client proposal needs to go out as a branded PDF by Friday, a workshop presentation needs slides by Wednesday, and the weekly content calendar is asking for three social images you have not made yet.

The old way of handling this involved opening Canva, fighting with templates that almost matched the brand, exporting, noticing the colors were off, and exporting again. Then it involved opening Google Slides and nudging text boxes around pixel by pixel until the deck looked passable rather than good.

This year to date, I have lived a compressed version of that problem. Within a few weeks I needed a full client proposal PDF for a custom build pitch, a presentation deck for an Enterprise Cayman workshop, and a fresh set of illustrations for a weekly newspaper column. Any one of those used to be an evening of production work. Together they would have eaten a week of nights, and the output still would have looked like it was made by someone who is not a designer, because it was.

The Stack

• ▸Claude Code, the AI coding agent that lives in your terminal
• ▸Google's Nano Banana models, accessed through the Gemini API
• ▸OpenAI's GPT Image models, accessed through the Images API
• ▸Custom skills: pdf, caydev-image, google-imagegen, and openai-imagegen commands wired into Claude Code

The combination turns natural language instructions into production-ready branded assets. No design tool, no template hunting, no export and reimport cycles.

A quick decoder on image model names

The naming landscape shifted fast over the past year, so here is where things stand as of this writing.

Google's image models all live under the Nano Banana family name, which started as a viral codename and became the official product line. Nano Banana 2 (Gemini 3.1 Flash Image) shipped in February and is the workhorse: fast, near-Pro quality, and around seven cents per image. Nano Banana Pro (Gemini 3 Pro Image) is the studio-grade option, slower and pricier, but it renders precise text inside images better than anything else I have used, which matters for banners and diagrams.

OpenAI's line is GPT Image. GPT Image 2 (gpt-image-2) shipped two weeks ago alongside ChatGPT Images 2.0 and is their new state of the art for generation and editing. The older GPT Image 1.5 still earns its keep for one specific reason: it supports transparent backgrounds, which GPT Image 2 currently does not, so logo and sticker work still routes there.

You do not need to memorize any of this, and that is the entire point of the next section.

The Upgrade

One command to PDF

/pdf "Create a 5-page client proposal for an AI consulting engagement.
Include: executive summary, scope of work, timeline, pricing, next steps.
Use Caydev branding."

Claude Code generates a complete PDF with the correct brand colors, professional typography, and content structured from the prompt. The brand palette is baked into the skill configuration, so there is no template to maintain and nothing to manually format. The custom build proposal I mentioned earlier went out exactly this way, from outline to sendable PDF in a single terminal session.

One command to social images

/caydev-image "Isometric illustration of a founder reviewing
a SaaS dashboard on a laptop, tropical office setting,
warm golden light, Caydev palette, 16:9, No text"

The caydev-image skill wraps the raw generators with brand rules: the Warm Tech aesthetic, Caribbean-inspired lighting, and the exact color palette are pre-configured. There is no prompt engineering trial and error each time, because the skill carries the style constraints, so I only describe the subject.

The routing is the real trick

Here is the part that makes multiple image providers practical rather than a chore. The skills choose the model for the job automatically.

The google-imagegen skill defaults to Nano Banana 2 for icons, social graphics, backgrounds, and batch work, then escalates to Nano Banana Pro on its own when the prompt involves precise typography, brand-critical hero assets, or complex multi-element composition like labeled infographics. The openai-imagegen skill defaults to GPT Image 2 and drops to GPT Image 1.5 when a task needs a transparent background or strict identity preservation on an edit.

Encoding those rules once means I stopped making model decisions entirely. I ask for the asset, and the routing logic spends my money sensibly: seven cents when fast is fine, ten when the text has to be crisp. When the newspaper column needed four editorial illustrations in a consistent style, that was four short commands, with the skill quietly picking the right model for each one.

The workflow in practice

A typical Monday now looks like this. At 9:00 I review the week's content calendar (after school drop off), which lands in my WhatsApp every Monday morning. By 9:15 I am generating the week's social images, three commands for three images. If a client proposal is due, that is one /pdf command while I eat my breakfast. By 9:45 I am back to actual work, meaning the product and client work that generates revenue rather than the production work that supports it.

The Impact

Before this system, document production and image creation consumed two to four hours a week, the branding drifted across assets depending on which template I started from, and the "I'll fix the formatting later" pile grew indefinitely.

After, the same output takes fifteen to thirty minutes a week, and every asset matches the palette automatically because the brand rules live in configuration rather than in my memory.

Speed is only half the win. When every PDF, social image, and slide deck shares the same visual language, your company stops looking like a one-person operation even when it is one. Consistency is what separates "founder who makes their own graphics" from "brand," and consistency is precisely what automation is good at.

Try It

The pieces you need:

1. Claude Code at claude.ai/code, with a Claude subscription
2. A Gemini API key from Google AI Studio for the Nano Banana models
3. An OpenAI API key for the GPT Image models
4. Custom skills, which are Claude Code configuration files that encode your prompts, brand palette, model routing rules, and output formats

The skills are the part worth copying. Start by writing down your brand rules once: colors as hex codes, the illustration style you want, the typography for documents. Wrap those in a skill file, add the model routing rules from the decoder section above, and every future asset inherits them automatically.

The key insight is that you do not need a design tool. You need a system that already knows your brand rules and applies them without being asked, and Claude Code skills are that system.

So here is the question worth sitting with: how many hours of your week are production work that a one-time configuration file could be doing instead?