Detailed Analysis
A common frustration among Claude Pro plan subscribers centers on token consumption within Projects, particularly when working with large document sets. As illustrated by this Reddit post, a user operating with seven PDFs ranging from 15 to 50 pages each finds that project-based workflows rapidly exhaust their monthly token allocation. The core concern is well-founded: Claude's Projects feature loads uploaded documents as persistent context, meaning every conversation within a project carries the overhead of those files. For users on the Pro tier primarily using Claude Sonnet, this architecture creates a tension between the convenience of persistent document context and the practical reality of hitting usage limits faster than anticipated.
The token consumption pattern described reflects a structural characteristic of how Claude handles project knowledge. Unlike standard chat sessions where context resets between conversations, Projects are designed to maintain continuity — a feature that is simultaneously their primary value proposition and their cost driver. When PDFs are stored in a project, Claude references that material as part of its system-level context, meaning the token cost is incurred whether or not the specific documents are directly relevant to a given query. This behavior is by design, prioritizing comprehensiveness and recall accuracy over token economy, which creates a challenging tradeoff for users working with large corpora on metered plans.
Several mitigation strategies exist, though their effectiveness varies depending on whether users are working via the web interface or through Claude Code. The most impactful approach involves maintaining a lean, structured summary document — analogous to a CLAUDE.md file used in Claude Code environments — that distills key facts, decisions, and reference points from the underlying PDFs rather than relying on the raw source documents to remain perpetually loaded. By front-loading critical information into a compact, curated reference and limiting direct document uploads to only those actively required for a given task, users can substantially reduce passive token overhead. Manual checkpointing — prompting Claude to summarize progress and key context before long sessions approach compaction thresholds — further protects against redundant re-processing of the same material.
The broader pattern here reflects a growing challenge across the AI assistant landscape: as these tools become more capable of handling complex, document-rich workflows, the infrastructure costs of that capability increasingly fall on end users through consumption-based limits. Anthropic's Projects feature was designed to address the persistent problem of context loss across sessions, but its implementation implicitly assumes users have either generous token budgets or relatively small knowledge bases. The friction experienced by this Reddit user is representative of a segment of the Pro user base that sits in an awkward middle ground — sophisticated enough to benefit from Projects' organizational features, but not operating at the enterprise scale that would justify an upgraded plan tier. This gap is likely to drive demand for finer-grained context controls, such as the ability to selectively activate specific project documents per conversation rather than loading the full project corpus by default.
Anthropic's trajectory in Claude Code — where features like `/compact preserve` and structured CLAUDE.md files give developers explicit tools to manage context economics — suggests the company is aware of this tension and is building toward more granular solutions. However, these controls remain largely confined to the developer-facing toolchain. Web-based Project users, who represent a substantial share of the Pro subscriber base, currently have fewer levers available. As token efficiency becomes an increasingly visible competitive dimension in the AI assistant market, closing this gap between developer and consumer context management tools is likely to become a meaningful product priority for Anthropic in the near term.
Read original article →