Skip to content

The Kitchen Sink Session

Mixing unrelated tasks in a single Claude Code session fills the context window with irrelevant history and degrades output quality.

Learn it hands-on with The Kitchen Sink Session guided lesson, which includes quizzes.

The problem

You keep one Claude Code session running all day and pile tasks onto it: review a PR, then start a feature, then debug a test failure. Each task leaves residue: file contents, command outputs, failed approaches, and off-topic reasoning. As the context fills, Claude starts making decisions shaped by stale information from earlier tasks.

The Claude Code best practices docs call this the "kitchen sink session" anti-pattern: context full of irrelevant information that degrades performance on the current task. The same source says LLM performance drops as the context fills, so the context window is the main resource to manage. Independent research by Chroma on context rot backs this up. Across 18 frontier models, performance varies widely with input length even on simple tasks, and irrelevant tokens hurt reliability more than length alone would predict.

Token costs reflect this directly. A session that runs through code review, feature development, and a debugging investigation builds up far more context than three focused sessions would. You pay for the noise.

What to do instead

Give each session a single objective. When you finish a task and move to something unrelated, start a new session, not a /clear.

Clear context within a session when you switch between loosely related tasks that still share background:

/clear

Resume a specific thread without re-entering context by hand:

claude --continue    # resume the most recent session
claude --resume      # choose from recent sessions

Claude Code saves conversations locally, so a new session does not lose prior work (Claude Code best practices). Use /rename to give sessions descriptive names (oauth-migration, debugging-memory-leak) so you can find them later with --resume.

For multi-step workflows with a clear dependency chain, use structured sub-agents. Each sub-agent runs in its own context window and reports back a summary, which keeps your main session clean.

Example

A developer spends the morning reviewing a pull request in Claude Code, then switches to scaffolding a new feature, then investigates a flaky test — all in the same session. By the third task, the context contains diff hunks, file reads, and unrelated error logs. Claude now has tens of thousands of tokens of history, most irrelevant. When asked what to name a new service class, Claude may anchor on naming patterns from the PR rather than the feature being built.

Correct approach, three focused sessions:

# Session 1: PR review only
claude "Review PR #412 — focus on auth logic"

# Session 2: Feature scaffolding only
claude "Scaffold the new billing service under src/billing/"

# Session 3: Debug flaky test only
claude "Investigate why tests/integration/auth_test.py fails intermittently"

Each session starts clean. Context stays low, costs stay low, and output quality is higher because Claude reasons only over relevant history.

When this backfires

Splitting sessions adds overhead. Claude re-reads CLAUDE.md and any shared context files on startup. If two tasks share a lot of background (a codebase you have already walked through, an architectural decision you set earlier), that re-loading cost can outweigh the noise you avoid with a clean start.

Auto-compaction also changes the calculus. Claude Code now compacts long conversations automatically as they approach context limits, summarizing decisions, file states, and patterns while dropping throwaway noise. For loosely coupled tasks where auto-compaction fires before quality drops, an explicit session split may not be needed. Use /compact by hand for finer control.

Split sessions remain the right call when: tasks are unrelated enough that shared history actively misleads (naming anchored to a prior PR, debugging instincts from a fixed bug); context is already full of failed approaches; or you are starting a review of code Claude itself just wrote in the same session.

Key Takeaways

  • One objective per session; use /clear between loosely related tasks in the same session.
  • Use claude --continue or claude --resume to pick up prior threads — session history is always available.
  • Token cost scales with context size; focused sessions are cheaper than long-running mixed sessions.
Established — multiple independent sources Last reviewed: 2026-06-13
Feedback