System prompt examples for real developer use cases

A system prompt is the part of an AI feature you write once and send every time. It sets the assistant’s role, what it may and may not do, and the shape of its answers, while the user’s message changes with each request. The prompts in this library were written for this site, not copied from any product, and each comes with notes on why it is built the way it is.

What the prompts have in common

Most share a structure that follows the providers’ own guidance. A one-sentence role comes first; Anthropic’s guide says even a single sentence of role focuses the model’s behaviour. Next comes the goal and the reason behind it, because explaining why a rule exists helps the model apply it to cases you didn’t list. Rules are concrete and observable (“escalate after two unsuccessful answers”, not “escalate when appropriate”), which is Google’s advice to give clear, specific instructions and constraints. The output format is shown rather than described. OpenAI suggests the same order for a developer message: identity, instructions, examples, context.

Two habits run through every prompt. They give the model an honest way out (“say the documents don’t cover it”, “write owner unclear”), which Anthropic lists first among ways to reduce hallucinations. And they treat anything the model reads, whether documents, tickets, emails or logs, as data rather than instructions.

Variables are per deployment, not per request

Placeholders such as {{company}} are filled once when you set the assistant up. Anything that changes per request, such as the user’s name, today’s date or retrieved documents, belongs in the user turn. That keeps the system prompt byte-for-byte identical across requests, which is what prompt caching needs: Anthropic’s cache only hits on a 100% identical prefix, and OpenAI recommends putting repeated content at the start of the prompt.

What a system prompt costs

The prompts here run from 258 to 522 tokens, 337 in the middle, counted with OpenAI’s o200k_base tokenizer. Sent with 10,000 requests a day, even a short prompt adds up, so each page prices it for a month on current models, with and without caching. Caching only applies above a minimum length: 512 tokens on Claude Sonnet 5.5, 512 tokens on Claude Haiku 5.5, 1,024 tokens on GPT-6.1 Sol and 4,096 tokens on Gemini 3.8 Flash. Below that, the prompt is billed in full every time; together with tool definitions, which sit in front of it in the cached prefix, it may cross the line.

Where it goes in each API

The text stays the same across providers; the field it goes in doesn’t.

# The same system prompt in each API (Python SDKs).
SYSTEM_PROMPT = open("support-agent.txt").read()  # fixed per deployment

# OpenAI Responses API
openai_client.responses.create(model=MODEL, instructions=SYSTEM_PROMPT, input=user_message)

# Anthropic Messages API (cached once it passes the model's minimum length)
anthropic_client.messages.create(
    model=MODEL, max_tokens=1024,
    system=[{"type": "text", "text": SYSTEM_PROMPT, "cache_control": {"type": "ephemeral"}}],
    messages=[{"role": "user", "content": user_message}],
)

# Gemini generateContent (google-genai SDK)
genai_client.models.generate_content(
    model=MODEL, contents=user_message,
    config=types.GenerateContentConfig(system_instruction=SYSTEM_PROMPT),
)

Every claim in the design notes was checked against the providers’ documentation on 2026-10-14. For a prompt you are writing from scratch, the prompt builder lays out the sections and shows the exact request; for one you are changing, the prompt diff shows what changed and what it costs.