Prompt Optimization

DEEVO

DEEVO, introduced in a paper titled Tournament of Prompts, evolves better prompts by having them compete. Candidate prompts face off in structured debates, an Elo rating tracks who tends to win, and the debate outcomes guide how new prompts are bred through crossover and mutation. Because it needs no ground-truth metric, only relative wins, DEEVO can optimize subjective tasks where there is no single correct answer to score against.

*These are learning tools, not prompts. They teach you to write your own, think of them as training wheels that guide you while you learn, then fall away once you can ride on your own.

The Core Insight

Rank Prompts by Who Wins

Optimizing prompts for subjective tasks is hard because there is no correct answer to measure against. What makes a persuasive essay, a tactful reply, or a compelling pitch is a matter of judgment, and a single numeric metric rarely captures it. Optimizers that need an absolute score stall here.

DEEVO sidesteps the need for an absolute score by using relative comparison. It borrows the idea behind chess rankings: you do not need to know a player's exact skill, only who beats whom. Candidate prompts are pitted against each other in structured debates, and an Elo rating rises or falls with each win or loss. Over many matchups, the Elo ratings become a reliable fitness signal without any ground-truth labels.

The debates do more than rank. Because each debate surfaces reasons why one prompt beat another, DEEVO uses that information to guide evolution, breeding new prompts through crossover and mutation that are informed by what actually won arguments. The loop repeats, and the population climbs toward prompts that consistently win on the qualities the task cares about.

Elo as a Label-Free Fitness Signal

An Elo rating needs only the outcome of matchups, not an absolute quality score. That makes it a natural fitness proxy for subjective tasks: DEEVO can rank and evolve prompts purely from which one wins a structured debate, so it works where no ground-truth metric exists.

The DEEVO Process

Debate, rate by Elo, breed from the winners, and repeat

1

Stage Structured Debates

Take pairs of candidate prompts and have their outputs compete in a structured debate, where the case for each is argued and judged. This produces a clear winner for each matchup without needing a numeric score.

Example

Two prompts for writing a cold outreach message go head to head; the debate judges which message would land better.

2

Update Elo Ratings

Each debate result adjusts the Elo ratings of the prompts involved, winners rise, losers fall. Across many matchups, the ratings converge into a stable ranking of the population's fitness.

Example

A prompt that keeps winning outreach debates climbs the rating, marking it as a strong parent for the next generation.

3

Breed New Prompts From Winners

Use crossover and mutation to create new prompts from the top-rated ones, informed by what the debates revealed about why they won. The evolution is guided by argument outcomes, not random change alone.

Example

The concrete opening from one winner is combined with the clear call to action from another to form a new candidate.

4

Repeat the Tournament

Run further rounds of debates, ratings, and breeding, so the population steadily improves. The best-rated prompt at the end is the optimized result, selected entirely through relative competition.

Example

After several rounds, the top prompt reliably beats the originals in debate, and it becomes the deployed prompt.

See the Difference

Needing an absolute score versus ranking by matchups

Absolute-Score Optimizer

Needs a Number

The optimizer requires a metric that assigns each output a quality score. On subjective tasks, defining such a metric is difficult and often unreliable.

Limit

Tasks whose quality is a matter of judgment resist a single score, so the optimizer has nothing dependable to climb.

Stalls when quality resists a single metric
VS

DEEVO

Ranks by Winning

Prompts debate head to head, and an Elo rating captures who tends to win. Only relative outcomes are needed, never an absolute quality number.

Reach

Works on subjective tasks with no ground truth, and the debates also inform how new prompts are bred.

Optimizes subjective tasks via relative wins

Practice Responsible AI

Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.

Most US states are actively legislating AI transparency and accountability. Critical thinking remains your strongest tool against misinformation.

DEEVO in Action

See how tournaments select and breed better prompts

Task

"Write a fundraising appeal." Persuasiveness is subjective, so there is no single correct output to score.

DEEVO Tournament

Debate: prompts whose appeals lead with a concrete story consistently beat those that open with statistics.

Elo: the story-first prompts rise in rating.

Breed: new candidates inherit the story-first opening and are tested again, climbing further.

Task

"Respond to an upset customer." The best reply depends on tone and tact, which resist a numeric score.

DEEVO Tournament

Debate: replies that acknowledge the frustration before offering a fix win against replies that jump straight to the solution.

Elo and breed: the acknowledge-first pattern rises and is bred into new prompts, producing consistently more diplomatic replies.

Two Winners

One prompt wins on clarity, another wins on warmth, each strong in a different debate dimension.

Crossover

Because the debates named why each won, DEEVO combines the clarity instruction from one with the warmth instruction from the other into a single new candidate.

Result: a prompt that carries both winning traits, which random mutation alone would have been unlikely to find.

When to Use DEEVO

Best for subjective tasks with no single correct answer

Perfect For

Subjective Quality Goals

Persuasiveness, tone, and style resist a numeric metric but are easy to judge in a head-to-head debate, which is exactly what DEEVO uses.

No Ground-Truth Labels

When there is no reference answer, relative wins still yield a stable Elo ranking, so DEEVO can optimize where scored optimizers cannot.

Combining Winning Traits

Debate-informed crossover can merge the strengths of different winners, which suits tasks where several qualities matter at once.

Creative and Marketing Copy

Content whose success is judged by impact rather than correctness fits a tournament that rewards the more compelling output.

Skip It When

A Clear Metric Exists

When accuracy or another objective score is available, an optimizer that targets it directly is simpler and more precise.

Debates Are Unreliable

If the judge cannot tell better from worse on the task, the Elo signal is noisy and the tournament can reward the wrong prompts.

Very Tight Budgets

Running many debates across a population costs model calls, so a strict compute limit may favor a lighter optimizer.

Use Cases

Where DEEVO delivers the most value

Marketing Copy

Evolve prompts for ads, emails, and pitches by rewarding the versions that win head-to-head on persuasiveness, not on a metric.

Customer Communication

Tune prompts for support and outreach replies where tone and tact decide quality, using debates to surface the better response.

Creative Generation

Improve prompts for stories, taglines, and concepts, where impact matters more than a correct answer, through competitive ranking.

Head-to-Head Evaluation

Rank a set of candidate prompts by relative strength when you can judge which output is better but not assign an absolute score.

Trait Combination

Merge the winning qualities of different prompts through debate-informed crossover, building candidates that are strong on several fronts.

Label-Free Optimization

Improve prompts for any task where relative preference is available but a ground-truth dataset is not.

Where DEEVO Fits

The tournament-based optimizer for subjective tasks

Debate Prompting Argue to Decide Debate improves answers
Pairwise Evaluation Compare Two Judge relative quality
DEEVO Tournament of Prompts Elo ranking, breed winners
GEPA Reflective Evolution Language diagnosis, Pareto
Chain These

DEEVO turns the ideas behind debate prompting and pairwise evaluation into a full optimizer, using Elo as the fitness signal. Reach for it when your task's quality is subjective and best judged by which output wins rather than by a score.

Let Prompts Compete

Explore how comparing outputs head to head can guide prompt design in the Prompt Builder, or see how DEEVO relates to the wider family of prompt optimizers.