What Fills Claude Code’s Context at Session Start, and How to Audit It with `/context`

A new Claude Code session starts with a fresh conversation context that already contains some information. Claude Code loads some project instructions and configuration before you send your first task. Claude adds to that context as it reads files, uses tools, and receives your messages.

A context window is the text and other information Claude can use while working in a session. Run /context immediately after launching a new session to see a live breakdown of its current usage. The view shows what uses context space, but it does not show every prompt detail or whether Claude followed an instruction.

What Claude Code loads before your first prompt

Claude Code may seem to remember an earlier session because it loads information saved in files. Project instructions and auto memory carry information between sessions; the previous conversation itself does not automatically become part of a genuinely new session. A resumed session is different, so use a new session when you want to measure a fresh start. The memory documentation describes how these files work.

Instruction files are a major source of startup context. Claude Code loads applicable CLAUDE.md files from relevant locations, including project and ancestor directories. Claude combines the instructions that apply to the current work. An @path import adds the contents of another file. Nested instructions generally load when Claude works with files in the corresponding subdirectory. AGENTS.md support also depends on the documented conditions.

Anthropic recommends keeping each CLAUDE.md under 200 lines because longer files use more context and can make instructions harder to follow. That is guidance, not a universal enforced limit. Imports count too: splitting instructions across files can improve organization, and imported content still uses context. Imports can nest up to four levels.

Auto memory is another source. It stores information Claude has learned, such as user preferences or project details, and loads the first 200 lines or 25 KB at session start. Claude reads other memory files when needed rather than loading every one at launch. Auto memory is on by default in local sessions, with documented exceptions such as some self-hosted and background sessions.

Claude Code also loads descriptions of available skills, which tell Claude what specialized capabilities it can call on. The full instructions for a skill load when Claude invokes it, so having several skills available does not mean every skill’s full contents are present from the start. MCP, a way to connect Claude Code to external tools and services, adds tool-related context. With tool search enabled, tool names and server instructions load at startup; Claude loads detailed tool definitions only when needed.

The startup context also includes system and tool information. Other configured content, such as an output style or text supplied through --append-system-prompt, can add to it. A SessionStart hook, which runs when a session begins, can also add context if it is configured to do so. Your setup determines which of these sources appear in a session. Anthropic does not publish the full system prompt, as noted in its settings documentation.

What changes after the session starts

A /context reading depends on when you take it. Before you send a task, it gives you a startup baseline. After you send a task, Claude’s prompt and the results of work add context. Reading files, running commands, using tools, and exchanging messages all contribute to the session’s contents.

Some information loads only when relevant. For example, subdirectory instructions apply when Claude works in that directory, and full skill instructions load when Claude invokes a skill. With MCP tool search enabled, detailed tool definitions also wait until needed. To compare startup usage with usage during work, take one reading immediately after launch and another after representative activity.

Audit a new session with /context

To compare sessions, keep the launch conditions consistent. Start in the directory you want to inspect and use the same host environment. Record which model Claude uses and the relevant configuration. Then:

1. Start a genuinely new session and run /context before entering a task. That reading gives you the clearest baseline for the session.
2. Run /context all to expand the view, including the memory files that loaded. The display varies by interface mode.
3. Run /memory to review or edit memory files. For settings sources, use /status as a complement: it identifies settings sources loaded for the session, while /context reports context usage.
4. Inspect likely contributors, especially instruction files and their @ imports. Also check auto memory, skills, MCP configuration, output style, appended prompt options, and any SessionStart hook.
5. Change one source at a time, then repeat the same baseline check. This makes it easier to see which change affected the reading.

The context-window walkthrough describes /context as a live view of usage by category, with optimization suggestions. The expanded view can help you check which memory files loaded. Treat the result as a useful audit view, not a complete transcript of everything Claude receives.

How to interpret the result

A large startup reading does not, by itself, indicate a bug. The result depends on the model, working directory, enabled integrations, configuration, and loaded instructions. Anthropic does not establish one universal startup-token total for all setups.

Use the categories and file details to identify which context sources you can change. If several instruction files or imports contribute to a large baseline, shorten or remove content Claude does not need at startup. If an integration adds context that the session does not need, adjust that configuration and measure again. Change one source at a time so you can attribute the difference.

Instructions tell Claude what to do, but they do not force it to comply. The /context view cannot show whether Claude followed them. When an action must be blocked independently of Claude’s decision, Anthropic recommends using a hook or another enforcement mechanism rather than relying on instruction text alone.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com