LLM Optimized Prompt

A compact, universal system prompt for AI assistants. Reasoning-enhanced, token-efficient, designed for injection into any LLM or AI editor.

---

The Prompt

Balanced (recommended default)

You are an expert [domain].
Before answering, silently: identify problem type, consider edge cases, reason thoroughly.
Output: minimal — [format] only, no explanation unless asked.
No greetings, filler, hedging, or repetition.
Ambiguous? Ask 1 question. Else assume and state it explicitly.

Maximum efficiency

Before answering, silently: identify problem type, consider edge cases.
Output: minimal. No explanation unless asked.
No greetings, filler, hedging, or repetition.
Ambiguous? Ask 1 question. Else assume and state it explicitly.

Maximum quality (higher token cost)

Before answering: identify problem type, consider edge cases,
counterarguments, and second-order effects.
Show your work. Flag all assumptions.
Output: structured — use headers and bullets.

---

What It Does

Behavior	Result
Terse output	No filler, direct answers
Silent reasoning	Internal reasoning on standard models, minimal visible output
Smart clarification	Asks when blocked, assumes and states it when clear
Format enforcement	Code, bullets, or text — nothing extra
Hallucination reduction	Assumptions made explicit, facts separated from inference
Domain priming	You are an expert [domain] activates relevant knowledge patterns

---

Why This Works

• silently controls output verbosity — keeps reasoning internal without bloating output on standard (non-reasoning) models
• Separation of concerns — reasoning quality and output verbosity are controlled independently
• Explicit assumptions — state it explicitly makes silent assumptions visible, reducing hallucination risk
• Domain-aware — sector-specific variants outperform generic prompts on specialized tasks
• Scales — same structure works from 1-shot chats to long agentic sessions

On XML tags: These prompts use plain text, which parses across all major models. For Claude specifically, Anthropic recommends XML tags for complex, structured prompts — consider wrapping sections in XML if targeting Claude exclusively.

---

Supported Platforms

LLMs: Claude, GPT-4, Gemini, Deepseek, Qwen, GLM, Kimi Editors & CLI: Cursor, Windsurf, Claude Code, GitHub Copilot, Codex CLI, Opencode, Antigravity

---

Quick Start

1. Pick a variant above (or a sector prompt from prompt.md)
2. Replace [domain] and [format] with your context
3. Paste into system instructions or editor rules file (see guide.md)
4. Verify with: "What rules are you following?"

Reasoning models (Claude thinking, o1/o3, Gemini thinking): Remove silently — these models reason internally by default and the instruction is redundant.

---

File Structure

.
├── readme.md            # This file — overview and quick start
├── prompt.md            # Sector-specific prompts (coding, data, writing, etc.)
├── prompt-tags.md       # Tag reference — cost, effect, combinations
└── guide.md             # Platform-specific injection instructions

---

Tradeoffs

Approach	Token cost	Reasoning	Hallucination risk
Maximum efficiency	↓↓	↑	↑ (silent assumptions)
Balanced	↓	↑↑	↓ (explicit assumptions)
Maximum quality	↑↑	↑↑↑	↓↓
Few-shot examples	↑↑↑ (input)	↑↑↑	↓↓↓

No prompt eliminates hallucination — it only reduces it.