Chat, Projects, or Cowork? Why Claude users are quitting in frustration – and how you can do better

Save up to 90% on tokens: The secret Claude feature that hardly anyone uses

"Message limit reached": The biggest mistake almost everyone makes at Claude AI

Anyone who uses the AI ​​Claude extensively knows this frustrating moment: right in the middle of their most important project, the message "Message limit reached" suddenly appears. The obvious reaction of many users is to cancel their subscription – but they're drawing the wrong conclusions. The problem is usually not Anthropic's supposedly strict limits, but rather how the tool is used. Those who treat Claude like a simple chatbot and explain complex blueprints, VOB clauses, or contracts to the AI ​​in every message will burn through their token budget in record time. This article explains why the AI ​​fundamentally operates without a state and how understanding the three fundamental levels – Chat, Projects, and Cowork – will unlock the platform's true potential. Learn how "Prompt Caching" can save you up to 90 percent of your tokens and transform the AI ​​into a highly productive, local assistant. An absolute must for efficient knowledge work – not only, but especially in the demanding construction industry.

The message limit isn't the problem. The problem is that almost no one knows what tool they're holding

Anyone who has worked with Claude in recent months knows the feeling: the progress bar runs dry, the message "Message limit reached" appears on the screen, and with it comes frustration. Especially when a calculation is halfway through or an expert opinion is in the middle of being written, the system seems to refuse to work at the worst possible moment. The obvious reaction of many users: cancel the subscription, switch back to ChatGPT, and shelve the whole AI issue for now.

What these users overlook is a technical connection that is crucial for the efficiency of any AI-supported work. Claude operates fundamentally without state. This means that every new message sent to the model triggers a complete reprocessing of the entire previous conversation history. A context of 50 messages costs not only the tokens of the new question on the 51st request, but the tokens of all 50 preceding messages combined. Token costs thus do not increase linearly, but quadratically with the length of the conversation – a mathematical principle that is ruthlessly exposed in a normal chat window.

Claude Pro offers around 45 messages within a rolling 5-hour window. That sounds generous at first. However, anyone trying to manage a complex construction project entirely within the standard chat window will find this limit becomes a structural problem: project history, VOB clauses, contract templates, and the construction schedule are explained to the AI ​​message by message. This consumes tokens at a rate that regularly surprises even experienced users. The Opus 4 model uses three to five times more tokens than the more affordable Sonnet model, which quickly leads to an empty daily token balance when performing complex document analyses.

The real flaw, therefore, does not lie in Anthropic's limitations. It lies in using a single, universal input channel for tasks that require fundamentally different infrastructures. Anyone who treats a hammer as a universal tool will find it utterly unsuitable for screwing.

Three levels, one AI: The architecture behind Claude

Anthropic designed Claude not as a monolithic chatbot system, but as a platform with three fundamentally different deployment modes, which differ significantly in setup effort, persistence, local access, and token efficiency. Understanding these three levels is the basic prerequisite for productive work – in construction as in any other industry.

The chat: Quick answers without memory

Chat mode is the most familiar way to access Claude and, for many users, the only mode they know. He is, quite literally, the clever stranger: highly competent in the moment, but without any prior knowledge or memory of previous conversations. Setup takes zero seconds. You simply open the browser window and start typing.

For a number of tasks, this immediacy is an advantage. A quick VOB question, brainstorming for an email to the client, rapidly reviewing a single contract clause – all of this is well-suited to chat mode. However, anyone who wants to manage entire construction projects, generate site diaries, and insist on VOB/B-compliant wording in the same window will pay a prohibitively high token price. Every message containing "please remember that we work according to VOB/B" is a pure waste of tokens.

The chat window also offers no lasting project context. Files can be uploaded manually—PDFs, Word documents—but these upload costs are incurred every time a conversation is started. Anyone who wants to use this mode for everything is actively burning tokens without any strategic benefit.

The projects: The persistent knowledge space in the browser

The Projects feature is the logical upgrade from the chat function and, at the same time, the most underrated productivity feature on the entire platform. A project is an isolated knowledge space within the browser: files are uploaded once—building permits, VOB text modules, contract templates, custom writing style instructions—and are permanently available. Claude never forgets this context. Every new conversation within the same project can access this knowledge base without requiring a single token to be used for retransmitting this foundation.

The underlying mechanism is Prompt Caching, which Anthropic has been systematically expanding since mid-2025. Only 10% of the normal input token costs are charged for the cached portion of the context. For a project with 200,000 tokens of uploaded reference documents, this means that from the second request onward, the system only pays for 20,000 tokens instead of 200,000. In practice, users report token savings of up to 90% compared to the regular chat mode, and latency—the waiting time for a response—is reduced by up to 85%.

This efficiency advantage is particularly relevant for the construction industry. Many tasks on construction sites are recurring: weekly site diaries, regular disruption reports, standardized change order reviews, and VOB-compliant correspondence with clients. All these tasks benefit maximally from the project context because the knowledge base remains stable, and only the most up-to-date information needs to be entered. Anyone who has set up this function will experience a qualitative change in their work: the AI ​​understands the context, the tone, and the contract structure. Explanations are no longer necessary.

Custom instructions within a project are also cached and, after their first use, cost only 10% of the original token expenditure. A 5,000-token instruction block that is active across 20 messages consumes approximately 84% fewer tokens with caching than the same instructions if they were manually inserted into each message.

Cowork: The desktop app as a local construction site assistant

Cowork is Claude's newest and, in terms of its implications, still least understood operating mode. Technically, it's an agent mode within the Claude desktop app for Windows and macOS, which makes Claude a tool fully integrated into the local work environment. The app is installed on the user's computer, runs in the system tray, and—with explicit user permission—can access local folders and files without requiring them to be manually uploaded.

This means that Claude reads specifications directly from the project folder on the desktop, analyzes Excel spreadsheets without going through the cloud, and can save new documents – revised reports, updated analyses, newly created spreadsheets – directly to the local hard drive. Switching between browser tabs and file managers is no longer necessary. The tedious copying process between the AI ​​interface and the user's computer is a thing of the past.

Cowork requires the Pro plan (€20/month) and technically runs on the Model Context Protocol (MCP), an open protocol that also allows Claude to connect to external systems such as CRM databases, project management software, or ERP systems. In the desktop app, "Extended Thinking"—the in-depth, step-by-step analysis of complex problems—is enabled by default. The model automatically determines how much computational effort a task warrants: simple questions are processed quickly, while complex problems are tackled with more in-depth planning.

For hours of deep work sessions on your own computer, coworking is the ideal format. There are no browser crashes, no lost conversation histories, no data loss from accidentally closing a tab. Work remains local, and the user retains complete control over file access.

The economic rationality of the three levels

Choosing the right deployment mode is not a matter of personal preference, but an economic decision with measurable consequences. In professional contexts, token consumption is a resource that must be managed strategically.

In chat mode without caching, each message costs the full token price. For a conversation with 100,000 tokens of context, this means that the same 100,000 tokens are charged again for every new message – a system that is only economically viable for short, context-free interactions. Someone who sends 10 requests daily in the chat window, each requiring 50,000 tokens of context, will pay 500,000 tokens daily – for context that never changes.

With active project caching, the cost for this consistent context drops to 10% of the original value. 500,000 token units become, in effect, 50,000 for the cached portion plus the actual new information. Anthropic states that for conversations with a 100,000-token document, Projects can achieve token savings of up to 90% and reduce latency by 79%. Even with a 10,000-token prompt, the cost reduction is still 86%.

For commercial API users, Anthropic has quantified this efficiency: Cached input tokens cost $0.30 per million tokens, while regular input tokens are charged at $3.00 per million tokens. One developer who switched their RAG system to prompt caching reported a reduction in their monthly costs from $8,000 to $800—a 90% saving simply by properly utilizing the available infrastructure.

These figures aren't directly transferable to professional users of the Pro plan, as they don't incur API fees per token. However, the principle is the same: the more efficiently the tokens are used, the more workload can be accommodated within the available message quota. Those who systematically cache their Claude context can increase their effective usage capacity many times over – without upgrading their subscription.

A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) – Platform & B2B solution | Xpert Consulting - Image: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors

Construction industry in digital transformation: AI as an answer to structural bottlenecks

The question of how to operate Claude correctly is not a trivial user guidance issue. It touches upon a deeper economic matter: The construction industry faces structural challenges for which digital tools like AI assistants are not a luxury solution, but an operational necessity.

Germany is currently facing a shortage of approximately 300,000 skilled workers in the construction sector alone. Construction costs are rising continuously, the complexity of projects is increasing, and regulatory pressure – VOB (German Construction Contract Procedures), mandatory BIM for public construction projects, EU AI regulations – is growing in parallel. In this environment, every hour a construction manager spends repeatedly typing the same information into an AI window is an hour lost. As trade publications have been emphasizing for years, the digitalization of the construction industry is no longer an optional addition – it is a matter of economic survival.

74% of construction companies in Europe already use AI in construction projects, particularly in design (48%) and planning (42%). 84% plan to increase their AI investments over the next five years. At the same time, Germany lags behind in European comparison: only 24% of German construction companies use ERP systems, compared to 45% in Belgium. This gap cannot be closed simply by increasing workload, but only through the smart use of technology.

According to current analyses, AI systems can accelerate planning phases in the construction industry by 40 to 60%. Construction progress monitoring, target/actual comparisons, risk forecasts – all these are tasks that AI performs more efficiently than manual processes. However, these efficiency gains only materialize if the AI ​​is equipped with the right knowledge context. A model that doesn't know which contract structure applies, which VOB variant has been agreed upon, and which project number should appear on the documents, produces work that needs to be corrected – and that costs more time than it saves.

Extended Thinking and Opus 4.6: When AI really thinks

With the introduction of Claude Opus 4.6, Anthropic has unlocked another dimension of professional AI use, particularly relevant for complex construction projects. The model features adaptive thinking: it independently decides how much planning depth and how many revision steps justify a task before providing an answer. Simple questions are answered directly. Complex analytical tasks—such as reviewing a multi-stage change order claim or assessing a component delay in a critical path analysis—are approached with multiple thought processes and systematic self-correction.

Opus 4.6 also introduces a 1-million-token context window for the first time in its beta version. For comparison, Claude's standard context window is 200,000 tokens – significantly larger than ChatGPT's 128,000 tokens. In practice, a 1-million-token window means that complete construction project documentation, including all plans, minutes, contracts, and email correspondence for a medium-sized project, can be processed in a single context without anything being missed. Anthropic describes this capability as a breakthrough for analyzing large datasets with consistent quality across the entire document.

Within Cowork, Extended Thinking is enabled by default. This is no coincidence: Cowork is explicitly designed for multi-hour, in-depth work sessions in which the model processes autonomous task chains – from analyzing a bill of quantities to creating a cost estimate and outputting a finished file to the local machine. Opus 4.6 achieves the highest score on Terminal-Bench 2.0, a benchmark for agent-based coding, and surpasses all other Frontier models on Humanity's Last Exam, a multidisciplinary reasoning test. For demanding knowledge work in the construction, legal, and financial sectors, it outperforms the next best model (GPT-5.2) by 144 Elo points.

The practical decision matrix: Which tool when?

The strategic use of Claude on the construction site follows a clear logic derived from the technical characteristics of the three modes. A simple rule of thumb helps with orientation: Short, context-free tasks belong in the chat. Recurring tasks with a stable context belong in a project. In-depth, file-intensive work belongs in Cowork.

Specifically, this means for a construction manager:

The morning begins with a quick VOB (German Construction Contract Procedures) question about a new contract clause – chat is the right approach. No setup time, no file management, immediate response. In the afternoon, the weekly construction log needs to be created – project is the right approach. The custom instructions for format and writing style have been set up once, the VOB-compliant text base is cached, only the day's events need to be entered. In the evening, a complex change order calculation based on a 200-page bill of quantities needs to be created – coworking on the user's own computer is the right approach. The AI ​​accesses the local document directly, analyzes the results thoroughly, and saves them in the project folder.

This distinction is not an academic exercise. It is the difference between an AI system that reaches its limits daily and causes frustration, and one that functions as a reliable colleague who knows the tools and the context.

Pricing model and economic reality of the Pro plan

Chat and projects are available in the browser. Cowork requires the desktop app, which is only accessible to paying users. The Pro plan costs €20 per month and includes approximately 45 messages per 5-hour window with Claude Sonnet, as well as significantly more limited usage with Opus. Those who work more intensively can upgrade to the Max plan: Max 5x offers five times the Pro capacity for $100 per month, and Max 20x offers twenty times the Pro capacity for $200.

The rolling 5-hour window differs conceptually from the daily reset logic of other providers. Messages sent five hours prior are removed from the quota and gradually replaced by new ones. This allows for more continuous usage throughout the day, but requires an awareness that intensive work periods will deplete the window more quickly than moderate use. Since August 2025, there have also been weekly caps affecting approximately 5% of users – an indication that Anthropic is striving for long-term capacity fairness.

Cross-platform usage – Claude.ai in the browser, Claude Desktop, Claude Code – draws from the same resource pool. Anyone writing code in Claude Code during the day and wanting to edit documents in the browser in the evening needs to keep this in mind. The good news: users who use their tokens efficiently through consistent project mode will get more out of the Pro plan in everyday use than a user with three times the limit who burns through them daily on chat walls.

Risks and limitations: What cowork can't do

Every technical solution has its limitations, and an honest analysis includes these. Cowork operates via the Model Context Protocol (MCP), which technically requires the desktop app to be open and the local MCP server to be active. Closing the laptop or exiting the app results in the loss of all active MCP connections. Scheduled autonomous tasks that require MCP access will fail if the computer is in sleep mode. This limits the ability to use Cowork as a fully autonomous background application.

Furthermore, there are currently compatibility issues: Locally configured MCP servers (installed via JSON configuration file) are not fully available in Cowork mode – only custom connectors configured via HTTP function reliably in Cowork. This means that anyone who has set up specific local integrations for construction software or ERP systems may need to reconfigure them for Cowork.

ARM64 support for Windows is still under development for Cowork, which may temporarily limit users of certain newer hardware. These are not fundamental shortcomings, but they illustrate that Cowork is still a relatively young technology that is constantly being developed.

The issue of data security also deserves attention. Granting Claude access to local project folders implicitly transfers trust in Anthropic's data protection architecture to sensitive construction and contract documents. For publicly commissioned construction projects bound by the German Construction Contract Procedures (VOB), this can have legal and data protection implications that should be reviewed before implementation.

Conclusion: Understanding the tool saves time and frustration

The message limit frustration experienced by many new Claude users is solvable. It's a symptom of a usage pattern, not a product flaw. Claude isn't a universal chatbot that can handle all tasks in one window. It's a platform with three different power modes that together form a complete AI workspace – from quick information lookups and a persistent project room to autonomous desktop work with local file access.

This differentiation is particularly valuable for the construction industry. For a sector suffering from a chronic shortage of skilled workers, increasing cost pressures, and growing regulatory complexity, properly implemented AI offers genuine relief. Not as a replacement for expertise, but as an intelligent assistant that understands this expertise, contextualizes it, and makes it applicable – week after week, document after document, project after project.

Anyone still managing projects exclusively through the chat window is not only missing out on 80% of the potential. They are actively wasting time and money – and not giving AI the chance to demonstrate its true capabilities.

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here or simply call me at +49 7348 4088 965. My email address is : [email protected]

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business