How to Prompt Google Gemini 4 Argon

Gemini 4 Argon thinks before it answers. You do not need tricks or long role-play. You need a clear goal, the right context, firm constraints, and a way to check the answer.

Tell Argon what done looks like, not how to think.

Goal

One or two sentences on what you want

Context

Only the files, data, and facts that matter

Constraints

Rules, limits, deadlines, and what to avoid

Output

Exact format, length, and sections

Verification

How it should check its work, such as citing pages or running tests

Goal: find every renewal risk in these 12 supplier contracts. Context: the attached contracts and our policy document. Constraints: use only the documents; flag anything unclear instead of guessing. Output: a table with supplier, renewal date, notice period, price change, risk level, and page reference, then a five-point summary. Verification: cite the page for every row and list any contract you could not read fully.

• Put large, unchanging material such as documents at the start of the prompt
• Put your specific question after the material
• Ask for page or section references
• Ask it to say when the answer is not in the documents

• Turn on tasks mode for long, multi-step work
• Say which connected apps it may use
• Say what needs your approval before acting
• Ask for progress updates

LOW

Classification, summarization, and extracting data

MEDIUM

The default. General coding, tool use, document Q&A, and video understanding

HIGH

Multi-file software engineering, formal proofs, architecture design, and finding vulnerabilities

Set the level in the API configuration, not in your prompt text. Test your task at all three levels.

• Use thinking_level, not thinking_budget, which returns an error
• Temperature, top_p, and top_k are ignored, so do not rely on them
• Use a response schema for structured output
• The default output cap is 65,536 tokens per streaming request
• Use long output only when one answer truly needs more, such as hard proofs or big coding tasks

Argon caches shared prompt beginnings of 6,144 tokens or more automatically. Keep repeated instructions and documents at the start of every request, and put the changing part at the end.

• Describe each tool clearly
• Prefer one tool call at a time in multi-step loops
• Return tool results in the same order as the calls
• Ask it to state its plan before acting
• Keep humans in charge of risky actions

• “Before answering, list what you would need to check.”
• “Give me your answer, then your confidence and what could make it wrong.”
• “Compare these options in a table, then recommend one with reasons.”
• “Write the tests first, then the code, then run the tests.”

• Pasting everything you have instead of what matters
• Vague goals such as “make this better”
• Asking for long outputs on simple tasks
• Using HIGH thinking for quick jobs
• Skipping verification