Skill Photo Album — AI skill for Claude Code

photo-album — a Claude Code skill for building a real printed photo book

Turn a large, messy pool of photos, videos and AI-edited images into a **print-ready vendor photo book** — with an auditable selection trail, a live album builder the user drives themselves, and gates that check the *printed pixels*, not just the plan.

Built with and for [Claude Code](https://claude.com/claude-code). The agent orchestrates; the deterministic steps are plain-Python scripts bundled here.

This skill is distilled from one complete run: a 30×30cm layflat book, 76 pages, ~480 pool images, 33 videos, two AI-chat conversations of edited images, 267 commits, 40 scripts and seven plan revisions. Every number quoted in the lessons is a measured number from that run — not a default to trust.

What it produces

Artifact	Description
ledger.json / .csv	One row per source image: id, source, dimensions, EXIF time, pHash, bytes, sha256
Inventory contact sheet	Offline HTML, browsable by source and by hour
selection.json	The re-culled pool: keep / keep+enhance / drop, with the judge verdicts behind each
Triage + duel records	Every vision judge's verdict file — the selection is fully auditable
pageplan.json	The user-ratified, spread-by-spread plan
The album builder	A single offline HTML file: live composer, the full pool, per-tile controls, presets, auto-fill, guides, overlays, page map, undo, JSON import/export
Print assets	Every planned image at print resolution, in the vendor's accepted format, with honest real-pixel DPI
Print-fidelity PDF	A proof rendered by the same engine that computes the final geometry
The book, in the vendor's editor	Assembled and reviewed — the user orders and pays

Quick start

/photo-album      (or just: "build me a photo book from this album link")

Then, in order:

1. Phase 0 — the agent establishes frozen contracts: title, product (trim size, page raster, p