Why this matters

The model does not see the raw repository by magic. Claude Code assembles a carefully shaped context before each turn, and this chapter follows that path from the first cached snapshot to the final prompt text.

Big picture first

Claude Code does not treat memory as one last add-on step. CLAUDE.md content enters during getUserContext(), memory mechanics can also enter during getSystemPrompt(), and override logic later chooses the final prompt body. That is the easiest way to understand this subtree: context and memory can feed the prompt in more than one place, but the final assembly rules still decide what the model sees.

How this part breaks down

The rest of the subtree keeps the teaching order readable, but the runtime can surface memory in more than one place:

1. context-snapshots-and-cached-inputs Start with the values Claude Code gathers before it even builds a prompt.
2. default-system-prompt-and-dynamic-sections See how the big default system prompt is built from cached and dynamic sections.
3. prompt-override-priority-and-final-assembly Follow the rules that decide whether default, custom, coordinator, or agent prompts win.
4. claude-md-and-memory-file-loading See how CLAUDE.md files and auto memory become model-visible instructions.
5. prompt-and-memory-data-structures Review the key types and object shapes before you revisit the deeper pages.