Detailed Analysis
A Reddit user's detailed bug report highlights a narrow but persistent issue with Claude's file upload system: certain PNG screenshots consistently fail to upload with a "Caricamento scaduto" (upload timed out) error, while the same visual content re-exported as JPEG uploads without issue. The user's troubleshooting is notably methodical, ruling out the usual suspects — file size (well under Claude's 30MB cap), network conditions, browser versus desktop client, and system-level factors — while isolating the failure to PNGs originating from one specific application, regardless of whether they're captured via macOS's native screenshot tool or the app's own image export feature. This points toward something in the PNG encoding itself (metadata, color profile embedding, chunk structure, or similar) rather than a generic size or network problem, since a basic macOS utility check (`sips -g all`) showed nothing abnormal on the surface.
The second phenomenon described — a chat session becoming "stuck" after repeated failed uploads, causing even previously-working images to start failing, followed by spontaneous recovery — suggests a session-state or rate-limiting bug distinct from the file-format issue. This is a common pattern in production AI chat systems where client-side retry logic, backend upload handlers, or conversation-state caching can enter degraded states after error conditions, sometimes only resolved by starting a fresh session. The fact that a new chat immediately cleared this symptom (while the original PNG-specific failure persisted across sessions) supports the user's own diagnosis that two separate bugs are likely at play: a transient session-level glitch and a more specific, reproducible file-handling defect.
This kind of granular, well-documented user report is valuable for identifying edge cases in multimodal AI products that handle diverse file inputs. Claude, like other chatbot platforms (ChatGPT, Gemini), must parse and process a wide range of image formats, color spaces, compression schemes, and embedded metadata across desktop, web, and mobile clients. Bugs tied to specific encoders or export pipelines — in this case, one particular macOS application's PNG output — are exactly the sort of long-tail issues that are hard to catch in standard QA testing but can meaningfully degrade user trust when they surface unpredictably in production. The user's workaround (converting to JPEG) is a reasonable stopgap but underscores a functional gap: users need lossless image fidelity for tasks like reading dense UI text or code screenshots, and being forced into lossy JPEG conversion is a real usability compromise.
More broadly, this thread reflects the growing pains of AI assistants as they scale into everyday professional and technical workflows where file upload reliability is now a core expectation rather than a bonus feature. As Anthropic and competitors push Claude deeper into coding, design, and document-heavy use cases, the infrastructure supporting file ingestion — upload pipelines, format validation, error handling, and session management — becomes as critical to the user experience as the underlying model's reasoning quality. Community-sourced bug reports like this one, especially when paired with a formal support ticket, often serve as an informal QA layer that surfaces platform-specific edge cases faster than internal testing alone, and they illustrate how even mature AI products continue to have rough edges in the "boring" plumbing beneath the chat interface.
Read original article →