brand-guideline-extraction

Brand Guideline Extraction

Reverse-engineer a brand's visual and verbal DNA from existing design assets, and codify it so a human designer or a generative pipeline can reproduce the look and feel.

The core insight: a brand guideline extracted from N assets is only as good as (a) how representative the corpus is, and (b) how rigorously you separate rules (what the brand always does) from variations (what it sometimes does) and exceptions (one-offs). The whole workflow is built around that separation.

Work through five phases in order. Phases 2's subagent fan-out is the expensive step; everything else is cheap.

Phase 1 — Asset inventory

Locate the assets. Usually the user provides a folder; if they name external sources (Meta Ad Library, a website, social profiles) gather what's accessible first, but never block on missing sources — work with what exists and note coverage gaps in the final document.

Build work/inventory.json, one entry per asset:

{"file": "path", "kind": "image|video|source_file", "width": 1080, "height": 1920,
 "format_class": "square|portrait|story|landscape", "content_type": "best guess: product-ad|announcement|seasonal|job-ad|...",
 "campaign_hint": "from filename/folder", "date_hint": "from filename if present"}

Folder names and filenames are metadata — Turkish, German, or any-language campaign names, dimensions in filenames ("1080x1920px"), dates, and platform markers ("META", "Google", "Pmax") all go into the inventory. Keep assets from different pipelines (e.g., Meta vs. Google vs. organic) distinguishable: they sometimes follow different sub-systems, and mixing them silently muddies the frequency analysis.

• Videos: extract 3 frames each (10%/50%/90%) with scripts/extract_frames.py — the 10% frame shows the hook, 50% the layout system, 90% the endcard/CTA convention.
• Design source files (.psd/.ai/.aep/.fig): don't skip them — run strings file.psd | grep -iE "font|typeface" | sort -u (and similar) to recover actual font names, which vision analysis can only approximate. Also record their layer/file naming if informative.
• Corpus size: 30–60 assets is the sweet spot. Under ~15, tell the user the guideline will be provisional and which content types are missing. Over ~80, sample proportionally across content types and dates rather than analyzing everything.