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.
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.
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
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.
Two prompts for writing a cold outreach message go head to head; the debate judges which message would land better.
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.
A prompt that keeps winning outreach debates climbs the rating, marking it as a strong parent for the next generation.
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.
The concrete opening from one winner is combined with the clear call to action from another to form a new candidate.
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.
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
The optimizer requires a metric that assigns each output a quality score. On subjective tasks, defining such a metric is difficult and often unreliable.
Tasks whose quality is a matter of judgment resist a single score, so the optimizer has nothing dependable to climb.
DEEVO
Prompts debate head to head, and an Elo rating captures who tends to win. Only relative outcomes are needed, never an absolute quality number.
Works on subjective tasks with no ground truth, and the debates also inform how new prompts are bred.
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
"Write a fundraising appeal." Persuasiveness is subjective, so there is no single correct output to score.
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.
"Respond to an upset customer." The best reply depends on tone and tact, which resist a numeric score.
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.
One prompt wins on clarity, another wins on warmth, each strong in a different debate dimension.
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
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.
When there is no reference answer, relative wins still yield a stable Elo ranking, so DEEVO can optimize where scored optimizers cannot.
Debate-informed crossover can merge the strengths of different winners, which suits tasks where several qualities matter at once.
Content whose success is judged by impact rather than correctness fits a tournament that rewards the more compelling output.
Skip It When
When accuracy or another objective score is available, an optimizer that targets it directly is simpler and more precise.
If the judge cannot tell better from worse on the task, the Elo signal is noisy and the tournament can reward the wrong prompts.
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
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.
Related Techniques
Explore complementary comparison and optimization techniques
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.