Detailed Analysis
A reproducible failure in Claude's Project Knowledge indexing feature has drawn public attention, with Max plan users reporting that Markdown files above approximately 15–40 KB trigger a permanent spinner during the indexing phase, silently preventing file content from being accessible in subsequent chats. The user documenting the issue conducted systematic testing by truncating their original document at multiple size thresholds, confirming that files below roughly 15 KB index successfully while anything larger stalls indefinitely — with no error message, no timeout notification, and no user-facing indication that anything has gone wrong. The affected files are standard UTF-8 encoded Markdown with no anomalous formatting, ruling out content corruption as a cause. The user also cross-referenced two GitHub issues — #25759 and #10841 — that document related but distinct failure modes: one in which Projects switch to RAG retrieval at just 2% of project capacity (well below Anthropic's documented threshold of "approaching or exceeding" the context window), and another in which files appear uploaded but remain silently inaccessible.
The significance of this failure lies in its incompatibility with the core use case Claude Projects was designed to serve. The product's value proposition rests on persistent, updatable knowledge bases — documents that grow over time and give Claude consistent context across sessions. A file size threshold that sits somewhere between 15 KB and 40 KB is, by practical standards, extremely low; a moderately active living document in Markdown can exceed 40 KB within weeks of regular use. More critically, because the failure is entirely silent, users have no way of knowing their knowledge base has stopped functioning until they notice, through degraded output quality, that Claude is no longer drawing on the expected context. The user's own experience — files that worked fine when smaller, then broke invisibly after routine updates — illustrates exactly how this failure mode evades detection.
The research context adds an important technical layer: indexing failures with Markdown files appear to be a broader, cross-platform pattern rather than a Claude-specific anomaly. Cursor, Obsidian, Sublime Text, and OpenAI's Custom GPTs have all exhibited analogous stalls or silent failures on .md files at scale, with causes ranging from explicit extension filtering in scanner configurations to plugin conflicts and parsing complexity. In the case of Claude specifically, research suggests that `.md` and `.markdown` extensions may be filtered out at the scanner level in the codebase indexing pipeline — meaning the failure is not a bug in the conventional sense but a design constraint that has not been surfaced to users through adequate documentation or interface feedback. The RAG fallback that silently activates in these cases may superficially preserve functionality while actually returning responses grounded in general model knowledge rather than the user's actual documents, creating a particularly insidious form of silent degradation.
The user's proposed remedies are technically reasonable and map directly onto the failure modes identified: raising the indexing threshold, surfacing explicit errors when limits are exceeded, providing retry controls for stuck jobs, and offering a user-facing toggle to disable RAG in favor of full context loading. Of these, the most immediately impactful would be a clear error state — because even if the threshold cannot be immediately raised, users who know their file has failed to index can take corrective action, such as converting to `.txt` format or splitting files. The current behavior, in which the system presents the appearance of successful operation while silently falling back to a degraded mode, is the most damaging aspect of the issue: it erodes trust in a feature that depends fundamentally on users being able to rely on what Claude knows.
Anthropic's trajectory toward agentic, long-context, and memory-augmented AI products makes reliable document indexing increasingly load-bearing infrastructure. As Claude is positioned for enterprise workflows, research applications, and extended personal productivity use cases — all of which involve documents that grow over time — silent indexing failures represent a category of reliability problem that undermines the credibility of the broader product direction. The pattern documented here, in which system limits are hit without user notification and failures manifest only as degraded output rather than explicit errors, is precisely the type of opacity that erodes user confidence in AI systems at the moment when those systems are being asked to take on higher-stakes, longer-horizon tasks.
Read original article →