SSkilvy
RULog inStart free

Controlling the model through settings, not just the prompt

Besides the prompt itself, you have levers to control the models behavior: a system prompt (sets the role and rules) and parameters (control the nature of the output). Mastering them turns "however it comes out" into predictable, tuned behavior.

System the models role and rules Temperature answer randomness/creativity Max tokens length limit Context message history
A system prompt sets the models behavior; parameters control the output.

The system prompt

A system prompt is an instruction setting the models role and rules for the whole conversation. "You are a support assistant for our bank. Answer briefly, politely, only about the banks products, do not give financial advice". It is separate from user messages and defines how the model behaves. This is the main tool for tuning behavior to the task.

Key parameters

  • Temperature: "creativity"/randomness. Low (near 0) — predictable, stable answers (for facts, structure). High — diverse, creative (for idea generation). For reliable tasks usually low.
  • Max tokens: a response length limit — control of cost and volume.
  • Others: parameters affecting word choice, repetitions (depend on the provider).
How do I choose temperature for different tasks?
Temperature controls the answers randomness. Guides: 1) Low (0–0.3): when predictability and accuracy are needed. For data extraction, classification, structured output, factual answers, code work. The model will answer stably and almost identically to one request — what reliable systems need. 2) Medium (0.5–0.7): a balance. For ordinary answers needing both correctness and naturalness. 3) High (0.8–1.0+): when diversity and creativity are needed. For idea brainstorming, text variants, creative generation where variability is useful. Practical rule: for most production tasks (where reliability and predictability matter) use LOW temperature. High — only where you consciously want diversity. A common beginner mistake is leaving high temperature for a task needing stability and being surprised why answers "jump". If you need the same answer format each time — set temperature low. Test it: run one request several times and see how much the answers differ — so you understand the parameters effect.

Context and history

An LLM API usually takes not one message but a history: system prompt + previous messages + new request. So the model "remembers" the conversation. But remember: a long context is more tokens (costlier and there is a limit). Managing what and how much you send in the context is an important skill for cost and quality. Do not drag the whole history if not needed.

Rule: a system prompt sets the models role and rules, parameters (temperature, max tokens) control the output. For reliable tasks — low temperature; manage the context for cost.

🧠 Which temperature for tasks needing predictability (data extraction, structure)?

← BackNext →