Optimize a prompt

Align prompts with your eval criteria using Prompt Learning - UI, SDK, Alyx, or Skills.

What is prompt optimization?

Prompt optimization (in Arize, primarily Prompt Learning) improves your Prompt Object using meta-prompting: an LLM reads your current prompt, evaluation feedback, and examples, then proposes a revised prompt that better matches your criteria. Instead of manual trial and error, the system runs a feedback loop - generate outputs, score them, feed the signals back, repeat - the same engine in the UI, SDK, Alyx, or Skills.

The prompt learning process

Why optimize a prompt?

Modern models are capable, but how you guide them drives reasoning, consistency, and accuracy. Most teams cannot hand-tune prompts at lab scale - which is why data-driven prompt optimization matters: you refine prompts from evaluations, natural-language feedback, and traces instead of guesswork, with the same engine in the UI or code.

In practice, this approach has improved coding accuracy on SWE-Bench by 10-15% and reasoning scores on Big Bench Hard by up to 10%.

What you get with Arize Prompt Learning:

Strong results

Proven improvements across key benchmarks including SWE-Bench for Claude Code and HotPotQA.

No-code or SDK

Optimize through the UI or directly via the SDK. Same optimization engine, flexible workflows.

Version control

All optimized prompts are automatically versioned in Prompt Hub for easy rollback, comparison, and tracking.

Experimentation

Run side-by-side experiments on optimized versions in the Playground to identify the highest-performing configuration.

Workflow

By Skills

Use the Arize skills plugin with the arize-prompt-optimization skill to optimize prompts from traces and experiments via the ax CLI. Try asking your agent:

• “Optimize this classifier prompt using the last week of eval feedback.”
• “Revise my support prompt to reduce verbosity while keeping escalation rules.”

By Alyx

1. Describe your goal

   Tell Alyx what to fix, for example “Optimize my prompt to reduce verbosity” or “Rewrite this to improve formatting consistency.”

2. Iterate in chat

   Ask for further adjustments until the result matches what you want.

3. Compare versions

   Review the changes in the Playground and save any version you are happy with to Prompt Hub.

By UI

Navigate to Prompt Hub, find your prompt, and click Optimize.

1. Name task and select dataset

   Enter a task name, then select the training dataset whose rows map to your prompt variables.

2. Select the experiment

   Select the experiment whose evaluation feedback will guide the optimizer.

3. Set the output column

   Choose the column that contains the model output you want the optimizer to improve.

4. Confirm feedback columns

   Confirm feedback columns. Evaluator results are auto-populated. Labels paired with explanations outperform binary labels alone.

5. Select meta-prompt model

   Choose the model that reasons over your examples and feedback to propose prompt updates.

6. Track and ship

   Monitor in Task Logs, compare versions in Prompt Hub, validate in the Playground, then tag the winner as Production.

For diffing and labels, see Save and version prompts.

See compare prompts side by side for the comparison workflow.

By Code

Prompt Learning via SDK

Automate prompt optimization programmatically using the Prompt Learning SDK, which iteratively refines prompts based on evaluation feedback and annotations.

The SDK supports advanced features, like built-in systems to run evaluators, and more flexibility to run multiple loops or define train/test splits.

Learn more about Prompt Learning via SDK →

Next up

When you have versions worth keeping, Save and version prompts in Prompt Hub so your team can track changes, experiment on different versions, and measure improvements over time.