3 Ways to Fix Claude Code's Context

Managing Context in Claude Code

Claude Code sessions may become slower and less reliable as their context fills. The video describes three ways to manage context: /compact, micro-compact, and /clear.

Why Context Builds Up

The 200,000-token context window includes the conversation, Claude’s responses, files it reads, and commands it runs. According to the narrator, performance can decline as context grows. Automatic compression begins at roughly 95% of the limit, and the session ends when the limit is reached.

The video also cites research on language models struggling to use information buried in long contexts.

Ways to Manage Context

• /compact manually summarizes the session at a chosen breakpoint. Custom instructions can guide what the summary preserves, such as project decisions. The video recommends using it after finishing one task and before moving to another.
• Micro-compact automatically cleans up older tool-call details while retaining conversation history. It runs in the background, so users have less control over what gets summarized and when.
• /clear resets the context for a clean start while retaining the project’s Markdown file, according to the video. It can help when switching to an unrelated task or when irrelevant context has accumulated. The narrator says earlier conversations remain accessible with the Up Arrow key.

Usage Tracking

The narrator mentions checking session usage with the cost command and refers to a Claude usage monitor for tracking usage.

Overall Guidance

Use /compact for deliberate summaries, let micro-compact handle routine cleanup, and use /clear when a fresh context is needed.

Speaker and Sources

The video features a single narrator from the BytecodeCrux channel. It refers to a Stanford study on long-context performance and Anthropic guidance on Claude Code and context management.