Workflow
Context Management
Keep context focused to preserve output quality.
Overview
Long conversations accumulate context that can degrade Claude's performance. As threads grow, responses become slower, less accurate, and sometimes contradictory. Managing context deliberately—keeping conversations focused, resetting when needed, and storing progress externally—maintains response quality throughout your session. Think of it like keeping your workspace organized instead of letting clutter pile up.
Use this workflow when conversations exceed 20-30 exchanges, switching between unrelated tasks, noticing degraded response quality, or working on multi-day projects that need persistent state.
Step-by-step process
Keep one task per thread
Start a new conversation for each distinct task. Don't mix refactoring, bug fixing, and feature work in a single thread—context switching confuses the model.
Store progress in scratch files
When working on multi-step tasks, save intermediate results to files. This externalizes state so you can reset context without losing progress.
Reset the session when drift starts
When you notice repetitive responses, contradictions, or declining quality, start fresh. Copy critical context to a new thread rather than continuing a degraded session.
Use explicit context boundaries
Signal major transitions within a thread. When changing topics, explicitly tell Claude you're moving to something new.
Example prompts
"I've noticed your recent responses are repeating previous suggestions and missing details we already discussed. Let's reset. Here's where we are: [summary of current state]. Let's continue from here with a fresh context."
"I'm working on a large refactoring over several days. Create a PROGRESS.md file tracking: completed modules, current module, remaining modules, known issues. Update this after each module so I can resume work tomorrow in a new thread."
"Our conversation hit 50 messages. Starting fresh. Previous context: We built a payment system with Stripe integration (see src/payments/*), added webhooks for payment events, tested happy path. Now: need to add refund functionality."
"Before we start this new feature, I want to clarify: previous discussions about the authentication system are not relevant here. We're working on a completely separate analytics module. Please don't reference auth patterns in your responses."
Expected outputs
Cleaner context
Focused conversations without accumulated irrelevant history, making it easier for Claude to produce accurate responses
Higher response quality
Maintained accuracy and relevance throughout long projects by preventing context degradation
Less confusion
Reduced contradictions and repetitions that occur when Claude tries to reconcile too much conflicting context
Persistent progress
External state storage ensures work isn't lost when resetting context, enabling multi-day projects
Best practices
Common pitfalls
Continuing degraded conversations too long
Solution: When you notice quality decline, reset immediately. Don't try to 'fix' a degraded context—start fresh with a summary.
Not preserving important context when resetting
Solution: Before resetting, save critical details to files: current state, decisions made, what's working, what's next. Reference these in the new thread.
Mixing multiple tasks in one thread
Solution: One thread = one task. If you need to work on something unrelated, start a new conversation. You can always return to previous threads.
Pro tips
Create explicit checkpoints
""Every 10 exchanges or after completing a phase, ask: 'Summarize where we are, what's working, and next steps' — Save this summary. It becomes your resume point.""
Use thread naming conventions
""Name threads descriptively: '[Feature] OAuth Integration - Phase 2' or '[Bug] Login validation error' — Makes context retrieval easier when reviewing past work.""
Set context budgets upfront
""Start complex tasks with: 'This is a multi-step project. After each phase, I'll create a summary in PROGRESS.md. Remind me to do this after completing each step.'""