01
What Prompt Optimization Is
Prompt programs, evaluators, traces, and why the unit of improvement is a workflow.
A short introduction to
Prompt Optimization
A book about turning prompts into measurable systems: GEPA, MIPROv2, evaluation design, agent traces, and the deployment habits that make optimized prompts reliable.
Josh Purtell
book draft
Prompt
Optimization
GEPA / MIPROv2 / evals / agents
promptoptbook.com
Abstract
A practical recipe for improving prompts.
Prompt optimization is the discipline of improving language-model behavior against an explicit task, dataset, judge, or environment. It begins where prompt engineering usually stops: with a repeatable measurement loop.
This book will introduce the core workflow for prompt optimization, from task definition and evaluation design through offline search with GEPA, instruction and demonstration optimization with MIPROv2, and deployment gates for production agents.
The focus is practical. The goal is to help readers build optimization loops that produce better prompts without confusing benchmark gains for reliable product behavior.
Chapters
Draft table of contents
02
Evaluation First
Datasets, rubrics, task containers, flaky metrics, and the line between signal and leaderboard noise.
03
GEPA
Reflective prompt evolution, candidate selection, edit proposals, and offline optimization loops.
04
MIPROv2
Instruction and demonstration search, Bayesian proposal policies, and budget-aware prompt tuning.
05
Agents
How prompt optimization changes for multi-step tools, browser work, code agents, and long-horizon tasks.
06
Deployment
Regression gates, monitoring, rollback, provenance, and when an optimized prompt is safe to ship.
Companions
Resources
Changelog
Status
June 2026
Initial public shell: abstract, chapter map, companion resources, and citation.
Next
Draft chapters on GEPA, MIPROv2, evaluator design, and production deployment.
Citation
Reference
@book{promptopt2026purtell,
author = {Josh Purtell},
title = {Prompt Optimization},
year = {2026},
publisher = {Online},
url = {https://promptoptbook.com}
}