A practical guide on saving Tokens (the basic billing unit for AI models) in Cursor (an AI-assisted coding tool) reveals a stark reality: 80% of users' consumed quotas are spent loading context (background information fed to the AI) rather than on model thinking. This means at least 80% of most people's AI coding bills are wasted money. We note that the crude, extensive era of AI programming is ending, and the era of precise accounting has arrived.

What this is

This widely-discussed long-form post summarizes a "stingy" methodology for AI programming, with the core logic being context control. The article points out three common wastes: First, letting the Agent (an autonomous mode that plans and executes tasks) aimlessly scan the entire project, reading dozens of irrelevant files just to change a single line of code. Second, using a single chat window until it goes dark, causing early, discarded historical information to be repeatedly billed. Third, using the most expensive Agent mode for the entire process to do light work like renaming variables. The correct approach is to replace global scans with precise references, decisively start new conversations after completing independent tasks, replace verbose natural language with structured instructions, and match different interaction modes—from lightweight to heavyweight—according to task difficulty.

Industry view

We believe this methodology marks the transition of AI programming from a "novelty toy" to a "production tool"—when tools are fully factored into costs, ROI (Return on Investment) becomes a metric that must be scrutinized. However, there are also risk warnings within the industry: over-focusing on saving Tokens will invisibly increase developers' cognitive load. People who should be focusing on business logic are forced to distract themselves calculating context lengths and selecting models; this is essentially shifting the platform's infrastructure efficiency problems onto the user. At the same time, excessively trimming background information may cause the AI to "miss the forest for the trees," producing fragmented code lacking a global perspective.

Impact on regular people

For enterprise IT: AI tool bills will not automatically optimize with productivity. Enterprises need to establish internal norms and train employees in "AI resource hygiene," otherwise cloud bills can easily spiral out of control.

For the individual workplace: Collaborating with AI is no longer simply "talking." The ability to use the most streamlined information to make the AI produce precise results—context management capability—is becoming a new hard skill in the workplace.

For the consumer market: The billing of various AI applications will become increasingly refined. Consumers, much like managing mobile data plans, will develop habits of monitoring AI quotas and proactively choosing lightweight modes.