he core issue facing AI users today is the compounding cost of context windows, where the model re-processes the entire history of a chat with every new input. The message you typed is the tiniest part of the overall call, as the entire conversation gets wrapped up and sent again from the top. This mechanism, while necessary for the model to 'remember' previous turns, leads to massive token waste, especially in long-running threads where the user's latest query is often just a rounding error compared to the historical data being re-processed.
Jones emphasizes that this is not a problem that model labs will solve, as they have an incentive to keep users engaged and consuming tokens. The more capable the tool, the more tools you hand it, the more material it puts in and the faster you hit the wall. Consequently, the responsibility for maintaining an efficient 'AI desk' falls entirely on the user. This requires a shift in behavior: from treating AI as a bottomless bucket for information to curating the context provided to the model.
Practical strategies include editing existing messages to correct errors, batching related questions, and strictly separating work stages into different threads. Starting a clean task when the job changes is a huge win and people seem to resist this. By carrying forward only the final artifact of a research stage rather than the entire history of rejected drafts and model reasoning, users can keep their context windows lean and avoid hitting usage limits.
For advanced users, Jones suggests moving beyond manual habits toward automated solutions. The 'Token Saver' skill automates basic cleanup, while the 'Ringer' multi-agent framework acts as a local intermediary. Ringer runs locally between your AI and the model provider, allowing it to return answers without a model call or select only useful passages to send through. These tools represent the next phase of AI productivity, where the goal is to enforce hard limits and ensure that every token spent contributes directly to the desired outcome.





