Steering long sessions

Planning & Scoping

• Use plan mode to have Claude research and propose a plan in read-only mode before execution
• Read through the plan thoroughly—a detailed plan reduces execution errors
• Iterate on the plan itself rather than hoping for the best during execution

Steering & Context Management

• Compact command: Summarizes conversation and frees up context window, but risks losing important details
• Add specific instructions after compact to control what gets summarized
• Specify which parts to focus on (e. g. , "focus on API changes only")
• Rewind: Double-tap escape on empty prompt to access checkpoint menu
• Restore code, conversation, or both
• Summarize from checkpoint (removes everything after it)
• Summarize up to checkpoint (removes everything before it)

Autonomous Work

• Goal: Set completion conditions describing what "done" looks like; Claude keeps working until evaluator confirms conditions are met
• Use /goal clear to cancel
• Conditions must be checkable from Claude's output (test results, type checker, etc. )
• Loop: Runs a prompt on fixed or self-paced intervals to pull external data (CI runs, deploys) and act on state changes
• Press escape to stop

Multi-Agent Coordination

• Work trees: Prevent multiple Claude sessions from conflicting over the same files
• Creates three independent file trees
• Clean work trees auto-remove on exit
• Use . worktreeinclude file to specify ignored files to copy into each work tree (e. g. , environment variables, local config)

Overall Strategy

• Scope work first, then steer
• Use goals when you can describe the end state better than the steps
• Run parallel work in separate work trees to avoid conflicts

Takeaways

• Use plan mode to research and propose a plan before execution, then iterate on the plan itself rather than during execution
• Use compact command with specific instructions to summarize conversation and free context, or use rewind to restore checkpoints and remove sections of conversation
• Set completion conditions with /goal to have Claude work autonomously until the end state is reached
• Use /loop to run prompts on intervals to pull external data and respond to state changes
• Use work trees to run multiple Claude sessions in parallel without file conflicts