Token efficiency

Cut waste. Keep signal. Every token costs money and context window space.

Use this as an operating constraint across the fundamentals skill family—not a substitute for those skills.

1. Document ingestion

• No raw PDFs/Word/rich files in chat. Convert to clean Markdown first.
• Large docs: extract the relevant section only. Never dump whole files.
• Use @file refs. Let the indexer handle it.

2. Conversation hygiene

• New chat per task. One topic, one thread.
• Past turn 12–15: start fresh; summarize prior work in 2–3 sentences.
• Batch related asks into one structured prompt.

3. MCP / plugin audit

Each MCP server injects tool schemas into every conversation. Idle servers are a silent tax.

Monthly audit: list enabled servers → count tools → disable unused or wrong-project servers → re-enable when needed.

Compression option: mcp-compressor wraps N tools into 2 (get_tool_schema + invoke_tool).

4. Model selection

• Deep reasoning, architecture, multi-file refactors: capable default model.
• Scoped edits, searches, shell: fast subagent when available.
• Formatting, rename, boilerplate: fast path or inline edit.

Audition ladder for cheap or unfamiliar models: untested → probation → proven.

• Audition on small, low-stakes tasks with a strong executed check (verified-delegation).
• Promote on evidence—first-try pass rate on your own tasks—not benchmarks or vibes.
• Route load-bearing work to proven models; keep a small slice of runs auditioning candidates.

5. Prompt signal-to-noise

Goal first. Constraints second. Context third. Lists over prose. @file over paste.

• Repetitive instructions → encode as rules/skills. Do not retype.
• Trust injected rules/skills. Do not restate them.
• Kill filler, hedging, and redundant politeness.

6. .cursorignore

Add .cursorignore at project root (gitignore syntax). Exclude build artifacts, locks, generated code, and heavy binaries so @codebase stays on real sources.

7. Always-apply rules audit

Keep alwaysApply only for truly universal standards. Prefer glob-scoped or description-triggered rules for file-type or stack-specific guidance.

8. Output token awareness

• Specify exact output format and length constraints.
• Request bullets, not prose, when that fits the task.
• Code gen: no narrating comments or redundant imports.

9. Task scoping

• Plan before coding when the path is unclear.
• Scope to a small file set; break big features; verify between steps.
• Wrong result? Revert and re-plan. Do not stack fix-on-fix.

10. Cache-friendly workflow

Keep rules/skills stable (they form the prompt prefix). Prefer stable guidance over frequent rewrites.

Anti-patterns

• Paste entire files when few lines matter
• One chat all day (drift past turn 12)
• MCP servers enabled “just in case”
• Restate context already in rules/skills
• alwaysApply on file-type-specific rules
• No .cursorignore in a large monorepo
• Manual @-tag of many files when search finds two or three

Fundamentals mappings

• Prompt SNR → dense questions in grill-me
• Document ingestion → glossary work in ubiquitous-language
• Task scoping → small steps in tdd-feedback-loops
• Index hygiene → module discovery in deep-modules-architecture
• Output awareness → precise interface deltas in interface-first-delegation
• Model selection → boss/worker routing in verified-delegation