← ClaudeAtlas

export-review-prooflisted

Export reviewable proof (screenshots, flows, logs) for Glassbox ground-truth review. Use this after implementing a change whose visible/behavioral/runtime effect should be reviewed — e.g. a UI change, a multi-step flow, or a computed result that needs evidence. Produces a version:2 manifest reviewable via `glassbox --ground-truth`.
brianwestphal/glassbox · ★ 34 · Code & Development · score 79
Install: claude install-skill brianwestphal/glassbox
# Export Review Proof for Glassbox You are working in a project that uses **Glassbox** for review. When you finish a change whose effect should be *seen*, don't just describe it — export **proof** that a reviewer (or a later AI session) can inspect. Glassbox is **consumer-only**: it runs no capturer. You write proof to disk; Glassbox reviews it. There are three kinds of proof. Pick by what you're proving. ## 1. Visual state — a screenshot For a change to a screen, component, or rendered output: - Capture the **actual** state as a PNG/JPEG (your test runner's screenshot API — Playwright `page.screenshot()`, etc.). One file per state worth reviewing. - Pair it with an **expected** image: a design spec, a reference render, or the previous run's actual (a regression baseline). The expected may live outside the repo (specs often aren't committed). - Glassbox scores the perceptual difference, hides identical pairs, and sorts most-different-first. ## 2. Behavior over time — a flow For a multi-step interaction (a journey, a sequence), choose one: - **One animated SVG.** Render the whole interaction to a single animated SVG (e.g. via domotion) and reference it as **one step's `actual`**. It animates live in the comparison view. Best when the *motion* is the point. - **Per-frame still steps.** One ordered PNG per step (`NN-step.png`), declared as a manifest **set**. Glassbox renders a per-step "Step k of N" navigator with per-step scoring. Best when each *state*