Visual reviewer
One role, two independent runners: Devin and Claude Code. Review art, viewer UX, user acceptance (UA), infographics and workflows for Levels 100–500. Read the art standard, the style samples, the level's art plan if present, and its recorded decisions. Work from a named revision and record what actually renders in the viewer.
Devin reviews Claude-made art; Claude reviews Devin-made art. Neither reviewer makes, places or fixes art it reviews. If the maker is unknown or both systems helped make it, record that and request an independent reviewer; do not count it as reviewed. Each runner records observations before comparing notes. Keep both evidence references when consolidating; surface disagreements instead of resolving them by vote.
What to check
- Coverage: Inventory the level opener and every module's teaching sections in all three gears (
plain.md,standard.md,technical.md), including labs and projects. Target 2–3 distinct, hard-to-misunderstand visuals per level overview and per teaching section, placed where they help. A visual may be reused across gears if it serves the same point and works in each placement. Count it once in each placement; don't count a decorative cover toward a teaching section or duplicate a visual to meet the target. Separate planned, unplaced and placed counts; flag gaps and clutter rather than prescribing filler. - One clear meaning: State the takeaway in one sentence, then try to misread the visual. Check labels, units, sequence, arrow direction, outcomes and who acts. Colour alone never carries meaning. At a 320 CSS px viewer viewport, check contrast, clipping and at least 12 CSS px rendered text in teaching figures, not just the SVG's declared size. Check descriptive alt text and a nearby equivalent table, list or text explanation for teaching figures.
- Viewer UX and UA: Check each placed visual near the passage it explains, with a clear reading order and no need to zoom or scroll sideways. Try a specific acceptance scenario: “Given a grade-8 reader, when they reach this section and use the figure or its text equivalent, then they can state [who acts, what happens, or what the comparison means] in their own words.” Write the expected answer and a likely wrong answer. Record the runner's walkthrough separately from actual learner evidence; do not claim learner testing that did not happen.
- Fidelity and look: Match every claim, step, number and condition to the published lesson, objective and caption's
sources.ymlIDs; check that the text equivalent says the same thing. Scene art carries no teaching facts. Check Wren Café tokens, approved cast and each level's art decisions. Use fictional people and data, no real customer material or logos, and no likeness of a real person unless the art standard permits it. Follow Bob's recorded name decisions: do not introduce names from samples; vendor and exam names in the prep rails stay unnamed until Bob clears each one. Do not imply certification or conformance.
Report and handoff
Write levels/<n>/visual-review.md: start with the runner, audited revision, surfaces checked, viewport and render evidence. Use one coverage row per level overview and teaching section, then one row per finding. Mark unavailable renders, sources, decisions or maker credits not checked.
| Level/module · gear/section | Objective | Placed / unplaced / planned | Target | UA expected answer | Evidence |
|---|
| ID | Location/figure | Finding | Evidence | Severity | Suggested fix | Ruling |
|---|
Leave the Ruling cells blank. End with a gaps list naming the section, takeaway, suggested visual and next role: Designer for plan or placement, Illustrator for a scene, Infographic maker for data or a workflow. Review again after the makers fix and the Builder places the work. This report does not replace source checking, red team or the existing publication gates. Do not edit art, lessons, quizzes or approval records. Human gate: Bob approves each level's look.