Decoding Claude: E03 - The Whiteboard That Keeps Everyone Honest
Imagine handing someone a sheet of paper with ten tasks on it. They read the list, keep it in their in-tray, and pile on top of it all the documents they need to accomplish the tasks. They might get the first two or three tasks right. Then the task list starts to become a distant memory. They then start to repeat work, skip steps, or simply improvise. That is what happens every time you give an AI a multi-step job.
If you are new to this blog series, start here. This builds on that.
The Agentic Whiteboard
An AI model's attention is divided among the content that exists inside its context window. It will pay more attention to the content relevant for the task at hand and less attention to things not immediately relevant. Therefore, the more deeply a task list is buried under a pile of documents or chat messages between us and the model, the vaguer the task list becomes for the AI.
Say you hire someone who can enter a flow state with each of their tasks. You have to somehow make the task list visible to them and remind them to look at it. So you put all the tasks on a whiteboard and it becomes easily visible to them. Then you slip in a sticky note every few tasks to remind them to look up at the board.
Claude's harness does the same in the form of a Todo Manager. It is like a whiteboard with all the tasks listed in sequence. Only one task on this board is allowed the status of "in progress". This forces the model to focus on one task at a time. If the model goes through 3+ rounds, a nudge is added to its context window to update its todo list. This nudge is enough to remind the LLM what the next few tasks in the list are.
The whiteboard keeps the agent on track when it comes to its task list. But what if the work requires pulling out an encyclopedia? Would you keep it on the desk after the research is done?
Next: Decoding Claude: E04 - Rent a Conference Room. Subscribe to get it when it lands.
New here? Start with Decoding Claude: How Agentic AI Actually Works — where the series began.
Source: S03 TodoWrite, learn-claude-code by shareAI-lab.
