Temperature

When an LLM generates text, it calculates a probability distribution over all possible next tokens. Temperature scales these probabilities before sampling. At temperature 0, the model always picks the highest-probability token (greedy decoding). At temperature 1.0, it samples according to the original probabilities. Above 1.0, the distribution flattens further, making rare tokens more likely. Think of it as a creativity dial: low temperature is precise and predictable; high temperature is creative and varied.

For code, lower temperatures (0-0.3) are usually better. Code has strict syntax rules and typically one correct implementation for a given specification. Higher temperatures can introduce syntax errors, incorrect API calls, or unusual patterns. However, slightly elevated temperatures (0.3-0.5) can be useful for brainstorming alternative implementations or creative problem-solving. Most AI coding tools set temperature automatically—Claude Code, for instance, uses optimized settings for different types of coding tasks.

// Temperature comparison (conceptual)
// Prompt: "Write a function to reverse a string"

// Temperature 0 — always produces the canonical answer:
function reverse(s: string): string {
  return s.split("").reverse().join("");
}

// Temperature 0.8 — might produce creative alternatives:
function reverse(s: string): string {
  let result = "";
  for (let i = s.length - 1; i >= 0; i--) {
    result += s[i];
  }
  return result;
}

You rarely need to adjust temperature when using AI coding tools like Claude Code—they handle it internally. If you are building your own AI-powered tool, start with temperature 0 for code generation and only increase it if you want more variation in outputs.