Claude Code Made Crystal Clear · chapter 2: Core Concepts: The Agentic Loop and the Context Window
The context gauge across one session
2026-09-01
Fill climbs with every file read and command result while response quality drifts down. Near the limit, compaction summarizes the conversation (a lossy drop); /clear resets the gauge entirely.
Below: the paragraph from the book that builds this idea, then the diagram itself (Figure 2.1), and a recap. About a minute of reading.
Long sessions do not fail suddenly; they sag. There are two compounding reasons. The first is quality: as the window fills with stale file versions, dead-end attempts, and old command output, the model must weigh all of it, and its attention spreads thin. The second is cost: on each request the full conversation so far is sent again, so every leftover token is paid for on every subsequent turn. As the costs guide puts it, "stale context wastes tokens on every subsequent message." Figure 2.1 shows both effects on one chart.
Recap
- The idea: Fill climbs with every file read and command result while response quality drifts down.
- The picture: Figure 2.1, from chapter 2 ("Core Concepts: The Agentic Loop and the Context Window") of Claude Code Made Crystal Clear.
- Go deeper: the chapter builds this step by step, with recipes and sources at the end.
This diagram is one of many in Claude Code Made Crystal Clear.
Every chapter opens with the gist, draws the hard ideas, and ends with recipes and sources.
Get the book
