Official Best Practices
Core Principle
The context window is your most important resource. It fills up fast, and performance degrades as it fills. Every strategy revolves around efficient context usage.
Six Key Strategies
1. Give Claude Verification
Single highest-leverage action. Provide tests, screenshots, expected outputs so Claude can self-verify.
Run the tests and ensure all pass before saying you're done.
If tests fail, fix the code — don't modify the tests.
Verification means include:
- Existing test suites
- Screenshots (paste images directly)
- Expected output examples
- Lint/type-check commands
2. Explore First, Then Plan, Then Code
Recommended workflow (press Ctrl+G to toggle Plan Mode):
Explore (Plan Mode) → Plan (Plan Mode) → Implement (Normal Mode) → Commit
- Explore: Let Claude read relevant code, understand existing architecture
- Plan: Claude drafts implementation plan, you review and adjust
- Implement: Switch to Normal Mode to execute
- Commit: Verify and commit
3. Provide Specific Context
- Use
@to reference files - Paste images (error screenshots, design mockups)
- Pipe input data:
cat logs.txt | claude - Specify constraints and reference patterns
4. Configure Your Environment
Configuration items by priority:
| Config | Purpose |
|---|---|
| CLAUDE.md | Project-level instructions (< 200 lines, concrete rules) |
| Permissions/sandbox | Security boundaries |
| CLI tools | gh, aws, gcloud, etc. |
| MCP servers | Connect external services |
| Hooks | Deterministic automation (not advisory) |
| Skills and subagents | Reusable specialized workflows |
| Plugins | Packaged and distributable extensions |
5. Manage Sessions
| Action | Shortcut/Command | Purpose |
|---|---|---|
| Stop | Esc |
Interrupt current operation |
| Rewind | Esc+Esc or /rewind |
Undo last step |
| Clear | /clear |
Clean context between unrelated tasks |
| Compact | /compact <instructions> |
Targeted context compaction |
| Side question | /btw |
Quick aside (never enters history) |
| Subagent research | Use Agent tool | Protect main context window |
6. Scale with Parallelism
- Non-interactive/CI:
claude -p "prompt" - Multiple sessions: Desktop app, web, Agent Teams
- Fan-out:
for file in $(cat files.txt); do claude -p "..." done
Common Anti-Patterns
❌ Kitchen Sink Session
Doing everything in one session. Context pollution degrades subsequent task quality.
Fix: /clear between unrelated tasks.
❌ Correcting Over and Over
Correcting the same error 2+ times. Claude may be stuck in a local loop.
Fix: After 2 failed corrections, /clear and rewrite the prompt.
❌ Over-specified CLAUDE.md
CLAUDE.md too long and detailed, burying critical rules.
Fix: Prune ruthlessly. Only keep what Claude can't infer from code.
❌ Trust-Then-Verify Gap
No verification means for Claude — just checking final results.
Fix: Always provide verification (tests, lint, type checks).
❌ Infinite Exploration
Claude keeps reading files but never starts actual work.
Fix: Scope investigations, or use subagents to isolate research.
Writer/Reviewer Pattern
An efficient collaboration pattern:
- Writer: Claude writes code
- Reviewer: Claude (or another subagent) reviews code
- Loop until review passes
This can be automated via skills or subagents.
Context Management Decision Guide
| When you need... | Use... |
|---|---|
| Project-level permanent instructions | CLAUDE.md |
| Path-specific differentiated rules | .claude/rules/ |
| Reusable specialized workflows | Skills |
| Delegate focused tasks | Subagents |
| Deterministic automation | Hooks |
| Connect external services | MCP servers |
| Package and distribute extensions | Plugins |
| Cross-session learning | Auto Memory |