Detailed Analysis
A user's attempt to build a custom "Skill" for Claude—designed to convert handwritten or Excel-based manufacturing production sheets into standardized Word documents—ran headlong into a hard resource ceiling on Anthropic's free tier. The user reports exhausting their entire token allowance after a single attempt in a fresh chat, despite starting from 0% prior usage, while running Sonnet 4.6 at "low" reasoning effort. In an effort to diagnose the problem, they published the full skill definition, offering a rare and detailed look at how a real-world, production-grade Claude Skill is architected: reference `.docx` templates consulted via Pandoc conversion to Markdown, a style guide for formatting rules, and a seven-step workflow spanning source-document parsing, PPE and equipment determination based on chemical hazard logic, protocol rewriting, and quality-control section generation.
The skill itself is a useful case study in why token consumption can spiral even under a "low" effort setting. Each invocation requires the model to read and reason over multiple reference files (the blank template, a fully worked example, and a style guide) before producing any output, and the instructions explicitly demand that the model consult these documents via Pandoc *before* every transformation step. That means every run reloads and reasons over potentially large document structures, ingredient-hazard logic, and multi-step protocol formatting rules—compounding token usage well beyond what a simple prompt would require. Image-based inputs (photos of handwritten sheets) add further overhead, since vision tokenization for images is markedly more expensive than text tokenization, and the skill's requirement to "carefully read the image, including all handwritten annotations" suggests significant visual reasoning work per run.
This incident highlights a broader tension in Anthropic's product strategy: Claude's Skills feature (part of the Claude Agent SDK and consumer/API tooling) is marketed as a way for professionals to encode complex, repeatable workflows—exactly the kind of enterprise use case Anthropic has been courting with Claude for regulated industries like manufacturing, chemical processing, and quality assurance. Yet the free tier's token ceiling appears fundamentally mismatched with the computational demands of sophisticated, document-heavy, multi-file agentic skills. A workflow that reads a Word template, a reference example, a style guide, and an image, then reasons through hazard classification and multi-step protocol generation, is closer to a professional/API workload than a casual chat interaction—yet it was being run entirely within the constraints designed for lightweight consumer usage.
More broadly, this reflects a recurring friction point across the AI industry as agentic, tool-using, file-referencing systems (Claude Skills, GPTs with retrieval, custom agents) move from novelty demos into real operational use. As users increasingly build genuinely useful automation—like transforming legacy manufacturing documentation into standardized, compliance-ready formats—they are discovering that the "skill" abstraction, while powerful, carries hidden computational costs that aren't always transparent from the user-facing interface. This creates pressure on providers like Anthropic to either offer clearer token budgeting/estimation tools before execution, tier pricing that better reflects agentic workload complexity, or more efficient caching of reference materials (e.g., persistent skill-file embeddings rather than re-reading full documents each session) so that free and low-tier users aren't blocked from testing genuinely production-relevant use cases after a single run.
Read original article →