Decoder. plain-English AI glossary

Prompt optimization

▲ Rising

Iteratively improving a prompt — through testing, measurement, and refinement — to boost model performance on a task.

Think of it like

Tweaking a recipe to get better results: adjust salt, temperature, and timing.

Example

Test two versions of a prompt with 100 examples. Version A scores 87%, version B scores 92%. Use version B and iterate further.

How it actually works

Optimization can be manual (prompt engineering) or automatic (using another model to generate/improve prompts, or using gradient-based search on soft prompts). Metrics matter: define what "better" means (accuracy, latency, cost). A/B testing with real users is the gold standard but requires scale.

For product teams

Lifts performance without retraining models. Quick wins for existing systems.

For engineers

Set up eval harness. Benchmark prompts systematically. Track versions and scores. Automate if possible (e.g., LLM-as-judge for scoring).

Automatic prompt engineering — Automatic prompt engineering is one approach.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome