The Hypothesis

Can AI auto-tag an entire image library from filenames, folder structure, and image analysis — making assets queryable by AI agents?

The Concept

AI agents that generate landing pages, emails, or ads need to know what images are available and what they represent. But most asset libraries are just folders of files with cryptic names. This experiment connects to an S3/R2 bucket, scans every asset, and auto-generates metadata tags using filename patterns, folder structure, and AI image analysis. A file at `/products/blue-widget/hero.jpg` gets tagged `product`, `blue-widget`, `hero`. The result: an asset library that AI agents can query semantically.

Asset library auto-tagging

The hypothesis

Can AI auto-tag an entire image library from filenames, folder structure, and image analysis — making assets queryable by AI agents?

The concept

AI agents that generate landing pages, emails, or ads need to know what images are available and what they represent. But most asset libraries are just folders of files with cryptic names. This experiment connects to an S3/R2 bucket, scans every asset, and auto-generates metadata tags using filename patterns, folder structure, and AI image analysis. A file at /products/blue-widget/hero.jpg gets tagged product, blue-widget, hero. The result: an asset library that AI agents can query semantically.

How it works

1. Connect to bucket — S3, R2, or similar object storage
2. Scan all assets — index files, folder paths, MIME types
3. Extract metadata from structure — folder names → categories, filenames → descriptors
4. AI image analysis — visual classification: product shot, lifestyle, icon, trust badge, headshot
5. Generate queryable tags — structured metadata: type, product, variant, usage context
6. Human review and enrich — grid UI for correcting and adding tags

From unstructured folder of files to a semantically queryable asset library.

What it explores

• Can folder structure and filename patterns reliably generate useful metadata?
• How accurate is AI image classification for product photography vs. lifestyle vs. icons?
• Do AI agents make better creative decisions when they can query assets semantically?
• What tag taxonomy works best for AI consumption vs. human browsing?
• How does the system handle new assets added after initial indexing?

What we found

• Filename and folder-based tagging was accurate for 70% of assets
• AI image analysis bumped accuracy to 88%
• Product shots and icons were classified reliably
• Lifestyle images and abstract graphics were the hardest — AI couldn’t determine intended usage context without human input
• AI agents with tagged assets produced significantly more relevant visual selections than agents choosing from untagged folders

Learnings

• Folder structure is the strongest signal — well-organised buckets produce better auto-tags than AI image analysis alone.

• A simple tag taxonomy (type, product, variant, context) covers 90% of agent needs — over-engineering the schema adds complexity without value.

• Periodic re-scanning for new assets is essential — the library needs to stay current without manual triggers.

• Human review via a grid UI is fast enough to be practical (~5 minutes for 200 assets) — full automation isn’t necessary if the review step is lightweight.

Where it goes next

This feeds directly into the self-optimising landing pages experiment. When the page agent can query a tagged asset library, it makes better visual choices. We’re testing whether the tagging system can also infer brand guidelines from visual patterns — not just categorise, but understand style.