# How to eliminate Composer token bloat and slow response times using scoped .mdc globs

**URL:** <https://forum.cursor.com/t/how-to-eliminate-composer-token-bloat-and-slow-response-times-using-scoped-mdc-globs/172340>\
**Category:** Guides\
**Tags:** performance, context, composer, rules\
**Created:** [September 19, 2026, 11:14am UTC](https://forum.cursor.com/t/how-to-eliminate-composer-token-bloat-and-slow-response-times-using-scoped-mdc-globs/172340 "2026-09-19T11:14:10Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![MemorySync](https://sea3.discourse-cdn.com/cursor1/user_avatar/forum.cursor.com/memorysync/32/121144_2.png) [@MemorySync](https://forum.cursor.com/u/MemorySync)\
**Post date:** [September 19, 2026, 11:14am UTC](https://forum.cursor.com/t/how-to-eliminate-composer-token-bloat-and-slow-response-times-using-scoped-mdc-globs/172340/1 "2026-09-19T11:14:10Z")

</div>

If your Cursor Composer is feeling sluggish or burning through token credits rapidly, check your `.cursor/rules/` configuration.

A common pitfall is marking multiple `.mdc` files with `alwaysApply: true`. When you have 4–5 rules set to `alwaysApply`, Cursor prefixes 3,000+ tokens of static instructions to every single Composer turn. By turn 10 of a session, prompt prefill latency skyrockets.

### Best Practice Fix:

1. **Use file-matching `globs` in your `.mdc` frontmatter rather than global application:**

```yaml
---
description: Rules for database migrations and queries
globs: ["prisma/ **", "migrations/**", " **/db/**"]
alwaysApply: false
---

```

2. **Keep only ONE minimal file (`core.mdc`) marked as `alwaysApply: true`:**  
Keep this file under 50 lines, focused solely on coding style and project-wide invariant commands (e.g. package manager to use, linter command).
3. **Query docs via MCP instead of pasting them into rules:**  
Don’t paste massive framework documentation or API references directly into `.cursorrules` or `.mdc` files. Instead, connect documentation MCP servers so Composer fetches only the relevant sections on-demand.

This keeps your prompt window lean and ensures Composer generates answers in seconds rather than stalling on massive context prefills.