The Agentic Post
Breaking
Perplexity vs ChatGPT Search vs Google AI Mode: Which AI Search Is Best in 2026?  ·  What Are Vision-Language-Action (VLA) Models? How Robots Learn to Follow Instructions  ·  How to Translate Documents With AI Accurately (and When Not To)  ·  MCP vs A2A: The Two AI Agent Protocols Explained in Plain English  ·  Is Grok Accurate for News? What Studies Show About AI Search Citations  ·  Is AI a Bubble? 7 Warning Signs Investors Are Watching in 2026  ·  
Home/Guides/AI Glossary
System Prompts Explained: How to Give AI Standing Instructions That Stick

System Prompts Explained: How to Give AI Standing Instructions That Stick

AI Glossary

What a system prompt is, how it differs from a normal prompt, and how to write custom instructions in ChatGPT, Claude and Gemini that actually change results.

A system prompt is a set of standing instructions an AI reads before every conversation, telling it who it is, how to behave and what rules to follow. Developers write them into apps, and everyday users can set their own through features like ChatGPT custom instructions, Claude project instructions or Gemini Gems. A good one saves you from repeating yourself every time.

What is the difference between a system prompt and a normal prompt?

A normal prompt is your request right now: “summarise this email.” A system prompt is the background briefing that applies to every request: “you are a helpful assistant for a small bakery; reply in British English; keep answers under 100 words.” Think of it as the job description, while each prompt is a task.

Where can I set a system prompt without coding?

  • ChatGPT: custom instructions in settings, or instructions inside a custom GPT or project; see custom GPTs explained
  • Claude: project instructions and personal preferences in settings
  • Gemini: Gems, which are saved assistants with your own instructions
  • Developers: the system message field in any AI API

What should a good system prompt include?

  • Role and audience: who the AI is helping and their level
  • Tone and format: length, language, bullet points or prose
  • Facts it should always know: your business name, products, policies
  • Rules: what to avoid, when to say “I do not know”
  • Examples: one short sample of an ideal answer, which works like few-shot prompting

Can you show an example system prompt?

“You help customers of Green Leaf Plant Shop. Answer in a friendly, plain way, under 120 words. Only recommend plants from the attached product list. If someone asks about a sick plant, ask for a photo and the watering routine before advising. If you are unsure, say so and suggest emailing our team. Never promise delivery dates.”

Notice it sets a role, tone, length, a source of truth, a process and clear limits.

Why does my AI ignore my instructions?

Common causes: the instructions are too long or contradict each other, the rule is buried in the middle, or a long chat has pushed it out of focus. Keep system prompts short and specific, put the most important rule first, start a new chat for new topics, and test with tricky questions. Our guide to common prompting mistakes covers more fixes.

Are system prompts secret or secure?

No. Users can sometimes trick an AI into revealing its system prompt, and text in emails or web pages can try to override it, an attack called prompt injection. Never put passwords or private data in a system prompt, and do not rely on it alone for safety. See AI agent security practices for building safely.

Up Next
Cheapest AI APIs Compared: GPT-6, Claude, Gemini, DeepSeek and Qwen Prices (November 2026)

Cheapest AI APIs Compared: GPT-6, Claude, Gemini, DeepSeek and Qwen Prices (November 2026)

Comparisons

AI API prices per million tokens compared for GPT-6 Sol and Luna, Claude 5.5, Gemini, DeepSeek, Grok, Mistral and Qwen, plus how to cut your AI bill.

The cheapest capable AI APIs in November 2026 are small, fast models like OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 at about 0.10 dollars per million input tokens, and Alibaba’s Qwen3.7 Flash at about 0.03 dollars. Flagship models cost 20 to 40 times more. For most apps, the smart move is to use cheap models for routine work and save premium ones for hard tasks.

How much do the main AI APIs cost?

Prices per million tokens (input / output), as listed by BenchLM on November 6, 2026:

  • Claude Opus 5.5: 4.00 / 20.00 dollars
  • GPT-6 Sol: 2.00 / 10.00
  • Claude Sonnet 5.5: 2.00 / 10.00
  • Gemini 4 Carbon: 2.00 / 10.00
  • Kimi K3.1 Pro: 2.00 / 10.00
  • Grok 4.7: 2.00 / 6.00
  • Gemini 3.8 Flash: 0.75 / 3.75
  • Mistral Large 4: 0.70 / 2.10
  • DeepSeek V4.1 Flash (October update): 0.30 / 1.20
  • Gemini 3.5 Flash-Lite: 0.30 / 2.50
  • GPT-6 Luna: 0.10 / 0.50
  • Claude Haiku 5.5: 0.10 / 0.50
  • Qwen3.7 Flash: 0.03 / 0.13

Prices change often and providers offer discounts, so confirm on each official pricing page before budgeting.

What is a token, and how much is a million?

A token is a chunk of text, roughly three-quarters of an English word. A million tokens is around 750,000 words, about ten novels. Input tokens are what you send, including instructions and documents; output tokens are what the model writes, and they usually cost several times more.

Which cheap AI model is best value?

  • Everyday chat, classification, summaries: GPT-6 Luna, Claude Haiku 5.5 or Gemini Flash-Lite
  • Huge volumes on a tight budget: Qwen3.7 Flash or DeepSeek V4.1 Flash, after checking data and compliance rules; see is DeepSeek safe
  • Good quality at mid price: Gemini 3.8 Flash or Mistral Large 4, which also has open weights
  • Hard reasoning and coding: GPT-6 Sol, Sonnet 5.5 or Opus 5.5; see best AI model for coding

How can I cut my AI API bill?

  • Route by difficulty: send simple requests to a cheap model and escalate only when needed
  • Use prompt caching: repeated instructions or documents are billed at a large discount by most providers
  • Use batch APIs: non-urgent jobs often cost about half
  • Shorten outputs: ask for concise answers, since output tokens cost more
  • Trim context: send only the relevant parts of long documents
  • Set spending limits per key and per agent run

Is the cheapest model always the best deal?

No. A cheap model that fails half the time and needs retries can cost more than a pricier one that gets it right first time, especially for agents that take many steps. Test on your real tasks and compare the cost per successful result, not just price per token. Our guides to how AI rankings work and how DeepSeek changed AI pricing explain the bigger picture.