Project Setup

Create a reference images folder with subfolders for style, composition, subject, brand, and output examples — give your research agent structured input.

The reference images folder is your visual knowledge base. Organize it: /references/style/ for aesthetic references, /references/composition/ for layout examples, /references/subject/ for topic-specific imagery, /references/brand/ for brand guidelines, and /references/output-examples/ for the specific images you want to replicate. The research agent scans these folders to build its style brief. More organized input = more precise output.

PDFs contain junk data that bloats context windows. Convert research papers to markdown before feeding them to your agents.

A PDF of the PaperBanana paper loaded directly would consume most of the context window with formatting artifacts, embedded images, and metadata junk. Converting to a clean markdown file strips all that noise, leaving only the text your agents actually need. Use any PDF-to-markdown converter, clean up the output, and save it as a .md file in your project. Your agents get the insights without the overhead.

Copy-paste the image generation API documentation into a local markdown file so agents can reference it without web access.

Go to the NanoBanana (or whichever image API you use) documentation page, click "Copy Page," paste it into a markdown file in your project. Now your generator agent has the full API reference locally — parameter names, example calls, response formats, error codes. No web search needed, no hallucinated API parameters. The agent reads the authoritative source and builds correct API calls every time.