Failure modes
Eight ways this goes wrong, each with its fix — including the one that only shows up after you export.
A field guide that only lists strengths is advertising. Eight things worth knowing before you stake a deadline on this, ordered by how often they actually bite.
| Failure | Severity | What happens | The fix |
|---|---|---|---|
| Export shock | High | The design is perfect in the browser and degrades on the way to PPTX or PDF. | Name the destination in the first prompt and design within its grammar. |
| Specifying, not briefing | High | You list contents, get a competent arrangement of them, and cannot tell whether it works. | Name the reader, the moment, and the one thing it must not fail at. |
| Re-prompting the whole thing | Medium | A small local problem triggers a full regeneration, losing everything that was already right. | Comment on the element. Reserve conversation for structural change. |
| Stale design system | High | Every output inherits it, so the whole team produces confidently off-brand work at speed. | Curate it like the shared asset it is; re-point Claude at the codebase after a rebrand. |
| Prototype mistaken for product | High | It looks real enough that someone senior assumes it ships next week. | Label it on share. One sentence, every time. |
| Exploration budget burn | Medium | A dozen directions consumes usage like a dozen directions, mid-deadline. | Explore wide early, narrow before the expensive revisions start. |
| Wrong surface entirely | Medium | A settled design gets rehearsed on a canvas, or an open question goes straight to Code. | Three questions before you open anything. |
| Preview drift | Low | A behaviour you built a team ritual on changes underneath you. | Treat capabilities as dated. Re-audition rather than trusting your memory of them. |
The three that are structural
Most of the table is operator error and yields to a habit. These three are properties of the product, and no amount of skill removes them.
A prototype is a performance
No production data, no auth, no backend. It looks real and is not — that is the point, and it is exactly what makes it cheap enough to throw away. The hazard is entirely in how it is received.
It authors; you direct
No layers panel, no pen tool, no hand-built component library. For pixel-level typographic control or illustration, this is a different instrument — a good one, but not that one.
It is a research preview
Anthropic Labs ships it with limits, and capabilities and integrations keep landing. Expect the product to shift under you.