Giving Opus 5.5 a simulated paint canvas: what a synthetic UX audit found
I create digital illustrations using Procreate and Photoshop, but I'm always looking at what's happening at the intersection of AI and traditional art techniques. I followed the AI art discourse closely during the Midjourney debates. I want to understand how other artists are using AI as a creative tool rather than a replacement.
Giving Opus 5.5 a simulated paint canvas is a web product that drew attention this week: Launch. We ran a synthetic UX audit on it, sending a persona suited to the site through its pages and recording what it hit.
What our persona hit
- product-grid section. Each card displays italic descriptive text ('A woman pauses over a newspaper in a sunlit cleaners' window...') positioned between the image and the technical caption ('Round 20 · GPT-6.1 Sol, max reasoning'), creating ambiguity about whether the italic text is user commentary, AI output, or curator annotation. The source of voice is inconsistent: 'Alice, who runs the project' is clearly third-person narration, while 'I kind of like it' reads as first-person, but both use the same typographic treatment with no attribution label. Heuristic: Gestalt, Proximity.
- content section. Watch affordances are inconsistent across the three sections visible in this viewport. The top three paintings use '▶ Watch it being painted' and '▶ Watch' links positioned below the painter's commentary text; the 'Left to choose' row embeds '▶ Watch' below each card title but above the model name; the bottom 'Same brief' section embeds watch links within individual card footers. The play-triangle icon and link color remain consistent, but the three different placement patterns mean users must re-scan each section to locate the video control. Heuristic: NN/g Heuristic #4, Consistency.
- product-grid section. Three cards include a '▶ Watch it being painted' affordance, but three others do not, with no visual signal explaining the availability pattern, users familiar with gallery sites expect either uniform media availability or a clear preview indicator (thumbnail, duration badge, or disabled state) that telegraphs which items have video. The inconsistency forces trial-and-error scanning to discover which paintings have process documentation. Heuristic: Jakob's Law.
- content section. The 'Left to choose, a jug and lemons' section presents six paintings in a single horizontal row with no grouping or visual breaks. All six share the same subject (jug with lemons), making differentiation dependent entirely on reading each card's metadata label. The compositional similarity compounds Miller's working-memory load: users must hold multiple near-identical images in mind to compare model choice, and the flat presentation offers no chunking strategy, no grouping by reasoning level, model family, or outcome variation. Heuristic: Gestalt, Proximity.
One thing done well
The caption 'Claude Opus 3.5, after Caspar David Friedrich. The model wrote every brushstroke as code and a simulation of oil paint carried them out, replayed here sped up. No image generator.' provides complete process context immediately adjacent to the painting, eliminating the need for users to navigate elsewhere or remember details from other pages. The technical disclosure positions the AI's role (code generation) and the human reference (Friedrich) in a single sentence with no ambiguity about what tool did what work.
This is a single-page heuristic pass by a synthetic persona, not user research; it can miss context and occasionally misreads elements.
This audit was run independently by Blinx; Giving Opus 5.5 a simulated paint canvas was not involved.
Curious what a persona suited to your site would hit?
Run this on your own site