Cursor Context Management

Understanding Context

What is Context?

Context = Information sent to AI with your prompt
- Selected code
- Open files
- @-mentioned files
- Conversation history
- .cursorrules
- Codebase index references

Context Window Limits

Model context limits (tokens):
- GPT-3.5 Turbo: ~16K tokens
- GPT-4: ~8K tokens (32K extended)
- GPT-4 Turbo: ~128K tokens
- Claude 3.5 Sonnet: ~200K tokens
- Claude 3 Opus: ~200K tokens

Rule of thumb: 1 token ≈ 4 characters

Context Sources

Automatic Context

Cursor automatically includes:
1. Selected code (highest priority)
2. Current file content
3. Recently edited code
4. Import/dependency information
5. .cursorrules content

Manual Context (@-mentions)

@filename.ts - Specific file
@folder/ - Folder contents
@codebase - Search indexed code
@docs - Documentation
@web - Web search results
@git - Git history
@terminal - Terminal output

Context Optimization Strategies

Prioritize Relevant Code

Most effective:
1. Select exactly the code you're asking about
2. @-mention directly related files
3. Reference patterns to follow

Less effective:
- Including entire files when only function needed
- @-mentioning unrelated files
- Large conversation history

Minimize Context Waste

Do:
- Select specific functions, not entire files
- Use @file only for relevant files
- Start new chat when topic changes
- Be concise in prompts

Don't:
- @-mention "just in case"
- Keep long conversation threads
- Include commented-out code
- Add unnecessary explanations

Context-Efficient Prompts

Inefficient:
"I have this code and it's not working and I need help
understanding what's wrong with it and how to fix it
because it throws an error when I run it"

Efficient:
"Fix TypeError: Cannot read 'map' of undefined
@api/users.ts line 45"

Managing Large Codebases

Selective Indexing

# .cursorignore - reduce indexed context
node_modules/
dist/
build/
*.log
*.lock
test/fixtures/
docs/

Strategic @codebase Queries

Specific:
"@codebase where is handleUserLogin defined?"

Too broad:
"@codebase show me all the code"

Using Summaries

For large files, ask for summary first:
"Summarize the main exports from @lib/utils.ts"

Then dive into specifics:
"Explain the dateFormatter function in detail"

Conversation Context

When to Start New Chat

Start new chat when:
- Switching to unrelated topic
- Context feels polluted
- AI seems confused
- Responses are slow/truncated
- Starting fresh task

Preserving Important Context

If switching topics but need context:
1. Copy important code/decisions
2. Start new chat
3. Paste relevant context
4. Continue with fresh window

Multi-Turn Optimization

Good flow:
Turn 1: "Create user authentication"
Turn 2: "Add password reset to auth"
Turn 3: "Add rate limiting to auth"

Bad flow:
Turn 1: "Create user authentication"
Turn 2: "How do I sort arrays in Python?"
Turn 3: "Back to auth - add reset"
(Context confused)

Context Indicators

Recognizing Context Issues

Signs of context overflow:
- Truncated responses
- AI "forgets" earlier discussion
- Slower responses
- Less accurate suggestions
- AI asks for information you provided

Solutions:
- Start new chat
- Reduce @-mentions
- Select less code
- Use larger context model

Monitoring Context Usage

Watch for:
- Response quality degradation
- Increased latency
- Incomplete code generation
- References to wrong files

Advanced Techniques

Layered Context

Layer 1: .cursorrules (always present)
Layer 2: @-mentions (specific files)
Layer 3: Selection (immediate focus)
Layer 4: Prompt (your question)

Optimize each layer:
- .cursorrules: Essential patterns only
- @-mentions: Only truly needed files
- Selection: Exact code in question
- Prompt: Clear, concise question

Context Priming

Start chat with context summary:
"Working on auth module. Using JWT, refresh tokens,
httpOnly cookies. Following patterns in @lib/auth.ts.
Need to add password reset."

Then ask specific questions:
"Create the reset token generation function"

Reference vs Include

Reference (efficient):
"Follow the pattern in UserService"

Include (uses context):
"@services/UserService.ts create similar for Products"

Use references when AI already has indexed context

Best Practices

Chat Session Management

1. One topic per chat session
2. Clear context between major changes
3. Summarize decisions before long gaps
4. Export important decisions to docs

@-mention Strategy

Essential: Files you're directly modifying
Helpful: Files with patterns to follow
Wasteful: Entire folders "for context"

Start minimal, add if AI needs more

Code Selection

Select precisely:
- The function with the issue
- The type definition needed
- The specific error location

Not:
- Entire files
- Multiple unrelated functions
- Large code blocks "just in case"