Gemini System Instructions — Setup, Examples, and Free Generator

Gemini calls the system prompt "system instructions" — same concept, different name. If you've been hunting for "Gemini system prompt" and finding nothing useful, that's why. This guide covers exactly how to set system instructions in the Gemini API, in Google AI Studio, and in Vertex AI, with copy-paste examples and a free generator that produces Gemini-ready instructions in two minutes.

System instructions in the Gemini API

In the generateContent API, system instructions go in the system_instruction parameter, separate from the messages array:

{
  "system_instruction": {
    "parts": [
      {"text": "You are a helpful coding assistant. Always explain your reasoning in code comments. Never use deprecated APIs."}
    ]
  },
  "contents": [
    {"role": "user", "parts": [{"text": "Write a Python function to parse JSON."}]}
  ]
}

The structure is similar to Anthropic's Claude API: a top-level system field, separate from the conversation messages. This separation lets the model treat the system instruction with stronger weight than user-turn content.

System instructions in Google AI Studio

Google AI Studio is the fastest way to test Gemini without writing code:

1. Go to aistudio.google.com
2. Click Create new prompt → Chat prompt
3. In the right-hand panel, find the System instructions field
4. Paste your system instruction
5. Type a user message and run

This is the easiest place to iterate. Edit the system instruction, re-run the same user message, see the difference. No API key, no code, no deployment cycle.

System instructions in Vertex AI

For enterprise apps, Vertex AI is the production endpoint. The system instructions parameter works the same way:

from vertexai.generative_models import GenerativeModel

model = GenerativeModel(
    model_name="gemini-2.5-pro",
    system_instruction="You are a senior data analyst. Always validate the data type before computing statistics. Always cite the column you're analyzing."
)

response = model.generate_content("Compute the average of the price column.")
print(response.text)

Vertex AI also supports prompt grounding (attaching documents that the model can cite), function calling, and structured outputs — all of which interact with the system instruction.

Recommended structure for Gemini system instructions

Gemini accepts plain prose, but a numbered or bulleted structure works better than dense paragraphs. Suggested template:

You are [ROLE].

Your primary goal is [GOAL].

Capabilities:
1. [Capability 1]
2. [Capability 2]
3. [Capability 3]

Rules:
- [Rule 1]
- [Rule 2]
- [Rule 3]

Output format:
- [Format requirement 1]
- [Format requirement 2]

If you don't know the answer, say "I don't have enough information for that" rather than guessing.

The free system prompt generator produces output in this exact structure when you pick a use case and toggle rules.

Length recommendations

Use case	System instruction length	Notes
Personal assistant	100-300 tokens	Identity + 3-5 rules
Customer chatbot	300-800 tokens	Identity + capabilities + 8-12 rules + output format
Coding assistant	400-1000 tokens	Add code style preferences and stack details
Domain expert	800-2000 tokens	Add domain knowledge, terminology, and few-shot examples
Complex agent	2000-5000 tokens	Add tool descriptions, planning rules, and handoff logic

Gemini's 1M and 2M token context windows mean you have room to be thorough. Don't compress aggressively if it costs clarity. The system instructions count against your input token budget on every request unless you use context caching.

Gemini-specific tips

• Use structured output. Gemini supports JSON schema enforcement. Combine a clear system instruction with a schema for reliable structured responses.
• Lean into the long context. Gemini handles long instructions well. If your prompt needs few-shot examples, paste them in.
• Be explicit about safety. Gemini's safety filters are stricter than ChatGPT's. If your use case is sensitive (legal, health, finance), state explicitly that the assistant is for information only.
• Test on Flash first. Gemini Flash is much cheaper than Pro. If your system instruction works on Flash, it will work better on Pro.
• Use context caching for stable instructions. Gemini supports caching the system instruction. After the first call, subsequent calls cost less.

Common Gemini system instruction mistakes

• Searching for "Gemini system prompt." The official term is "system instructions." Google's docs use this exact phrase.
• Putting instructions in the first user message. The dedicated system_instruction parameter has stronger behavioral weight.
• Mixing capabilities and constraints. Separate "what you can do" from "what you must never do" — Gemini follows clean separation more reliably.
• Triggering safety filters. If you're building a healthcare or legal assistant, frame the system instruction around providing information, not advice. Otherwise the safety layer may block responses.