Audit Cursor Context

Audit Cursor Context

Investigate what is consuming the always-on portion of the Cursor context window in this workspace and recommend specific cuts. Always-on layers (system prompt + tools + rules + skills + MCP) load before the user types — every token there is a token unavailable to the conversation.

Reference

Companion research doc: docs/research/cursor-context-budget-audit.md (findings, recommendations, May 2026 best practices).

Methodology

Run the seven steps below in order. Use the table at the end as the deliverable.

Step 1: Read the user's reported context-bar numbers

Ask the user to share their Cursor context bar values, or read them from a screenshot if attached. The eight layers to capture:

Step 2: Inventory .cursor/rules/

for f in .cursor/rules/*.mdc; do
  echo "=== $f ==="
  head -5 "$f"
  wc -l "$f"
done

Flag every rule with alwaysApply: true. Healthy baseline: ≤3 always-applied rules totalling <3K tokens (≈12KB).

Step 3: Measure AGENTS.md overhead

wc -c AGENTS.md   # bytes
wc -l AGENTS.md   # lines
echo "≈$(($(wc -c < AGENTS.md) / 4)) tokens"   # 3.7 chars/token rule of thumb

If AGENTS.md exceeds 8K tokens (≈32KB), it is too large. The ceiling matters because AGENTS.md is always loaded.

Step 4: Inventory skills

ls .agent/skills/
# personal skills
ls ~/.agent/skills/ 2>/dev/null
ls ~/.agent/skills-cursor/ 2>/dev/null

Skills are progressive-disclosure — only their frontmatter (name + description) is loaded until invoked. Each costs ≈50–250 tokens. Do not recommend deleting skills to save context; recommend disable-model-invocation: true on skills that should only run when explicitly typed (/skill-name).

Step 5: Inventory MCP servers

cat .cursor/mcp.json 2>/dev/null
ls ~/.cursor/projects/*/mcps/ 2>/dev/null

For each server, count tools (each MCP tool descriptor lives at mcps/<server>/tools/*.json). Cursor enforces a 40-tool ceiling. If the workspace ships >40 tools, Cursor silently drops the overflow.

Step 6: Map each context-bar number to its source

Use these rough conversions to attribute bytes → tokens (English prose, ~3.7 chars/token):

Step 7: Produce the deliverable

Output a single markdown table the user can act on:

| Layer  | Reported | Largest contributor                                               | Recommended action                                                   | Reclaimed (est.) |
| ------ | -------- | ----------------------------------------------------------------- | -------------------------------------------------------------------- | ---------------- |
| Rules  | 59.3K    | `AGENTS.md` (147KB ≈ 37K tokens)                                  | Split into lean core + glob-scoped `.cursor/rules/learned-*.mdc`     | ≈30K             |
| Tools  | 22.4K    | `cursor-ide-browser` MCP server (~20 tools, verbose descriptions) | Trim per-tool descriptions to ≤50 words; disable unused tools        | ≈8K              |
| Skills | 4.4K     | 30 skills × frontmatter                                           | Add `disable-model-invocation: true` to skills only invoked manually | ≈0.5K            |

Decision rules

• Rules > 20K tokens → audit AGENTS.md size first. Anything over 8K tokens in a single always-loaded file is the prime suspect.
• Tools > 15K tokens → enumerate MCP servers; suspect the one with the longest serverUseInstructions and most verbs.
• Skills > 10K tokens → unusually high; skills are normally cheap. Re-check whether any skill is being inlined instead of frontmatter-only.
• MCP > 10K tokens → trim serverUseInstructions blocks (move detailed playbooks into a skill).
• Always recommend layering a lean AGENTS.md (architectural core, ≤8K tokens) on top of glob-scoped .cursor/rules/*.mdc for long-tail facts. Never recommend deleting AGENTS.md outright — it is portable across Claude Code / Copilot / Cursor.
• Never recommend deleting skills, only marking them disable-model-invocation: true.
• Never recommend disabling Cursor's dynamic context discovery (tool-output-to-file, summarization-with-history-search) — those are wins, not costs.

Refactoring patterns

Pattern A: Split a fat AGENTS.md paragraph into a scoped rule

Move the long-tail "Learned Workspace Facts" / "Learned User Preferences" paragraphs into focused, glob-scoped .cursor/rules/*.mdc files:

---
description: Graphics-pipeline facts for the CAD viewer (Three.js, WebGPU, TSL)
globs:
  - 'apps/ui/app/components/geometry/**'
  - 'apps/ui/app/machines/graphics.machine.ts'
alwaysApply: false
---

The agent loads the file only when the user is touching matching paths. Token cost when not touched: zero.

Pattern B: Convert a verbose MCP tool description to a skill

If an MCP server ships paragraph-length serverUseInstructions, move the detailed playbook into a .agent/skills/<name>/SKILL.md and shrink the server-level instructions to ≤50 words pointing at the skill name. Skill bodies load only when the skill is invoked.

Pattern C: Convert an alwaysApply: true rule to glob-scoped

Rewrite the rule's frontmatter:

# Before
alwaysApply: true

# After
globs:
  - "**/*.ts"
  - "**/*.tsx"
alwaysApply: false

Audit the body to confirm every guideline is genuinely scoped to those file types. If one guideline is universal (e.g. "use pnpm"), keep that one as a separate small alwaysApply: true rule (≤300 tokens) and scope the rest.

Anti-patterns

• Don't delete AGENTS.md — split it instead. Other agents (Claude Code, Copilot) read it.
• Don't delete skills to reclaim Skills-line tokens — the line is already small (frontmatter only).
• Don't measure tokens with wc -w; use bytes / 3.7 (English prose) or bytes / 4 (mixed code) as the conversion.
• Don't recommend running every audit through ctxaudit if the workspace already has the seven-step shell methodology above; the CLI is a convenience, not a requirement.

External tooling (optional)

sanjeed5/ctxaudit is a Python CLI that automates Steps 2–6 across user-level and project-level Cursor configuration. Use it as a sanity check after manual cuts:

pipx run ctxaudit

Output checklist

When this skill finishes, the user should have:

• A table mapping each context-bar layer to its filesystem origin
• An explicit "largest contributor" call-out per layer
• A prioritized list of cuts with estimated tokens reclaimed
• No recommendation to disable dynamic context discovery
• No recommendation to delete AGENTS.md outright
• No recommendation to delete skills

Source: taucad/tau — distributed by TomeVault.