Improve Cursor Indexing by Pruning Files and Selecting Optimal Models

Cursor indexing speed and accuracy increase when you remove unnecessary files, choose the right model, and use strategic prompts.

Indexing an entire file system creates noise and hallucinations. To stop this, use a .cursorignore file to exclude node_modules, .git, build directories, *.log, *.json, *.csv, *.sql, package-lock.json, and yarn.lock. This specific configuration reduces instances of the model failing to find obvious functions by 30%.

Retrieval accuracy varies by model. Claude 3.5 Sonnet excels at tracing long-range dependencies and identifying call stacks on the first attempt. GPT-4o is faster but oversimplifies, and Gemini 1.5 Pro is better for ad-hoc tasks than full integration.

For better efficiency, combine @Codebase for discovery with @File or open tabs for pinning. Using Composer mode (Cmd+I) with a file map, such as @auth.ts @middleware.ts @Codebase fix the session timeout bug, raises one-shot fix success rates from 40% to 80%.

For developers creating specs and use cases via OpenAI Codex, Project Assist provides an AI assistant called Herbs Assist. It streamlines workflows by producing Herbs spec files and natural language specifications. To start, run herbs assist in the terminal. If OpenAI Codex environment variables are missing, the assistant will prompt you to set them. When the tool asks, provide a use case name and the target folder or group. Herbs Assist then generates a natural language specification; if the suggestion is inaccurate, edit it before saving and moving to the next step.