In an evolutionary prompt search, what do crossover and mutation actually do?

## Population instead of an incumbent Hill-climbing carries one prompt and asks "can I do better than this?". Evolutionary search carries a population — say twenty prompts — and asks "which of these deserve to be parents?". The extra state is the point: several distinct lineages explore different regions of prompt space at once, and good ideas from different lineages can be merged rather than competing until one wins. ## Representing a prompt so operators are meaningful Operators only work if the representation has structure. A prompt is rarely one undifferentiated blob; it usually decomposes into a role or persona line, the task instruction, constraints and edge-case rules, few-shot exemplars, and an output-format spec. Treating those as fields gives crossover natural cut points and gives mutation a target to pick. Splicing two prompts at a random character offset produces garbage; splicing at section boundaries produces a coherent child. ## Mutation operators Mutation is a unary operator: one parent in, one variant out. Typical operators are rewording an instruction while preserving intent, tightening or loosening a constraint, adding an explicit negative rule, replacing one exemplar, reordering sections, or changing the requested output format. Two knobs matter. **Rate** — how many sections are perturbed per child — controls step size; too low and the population stagnates, too high and children stop resembling their fit parents. **Strength** — small rewording versus wholesale replacement of a section — controls whether you are exploiting a good region or jumping to a new one. Mixed operator sets, with mostly small edits and occasional large jumps, tend to behave better than any single setting. ## Crossover operators Crossover is binary: two parents in, one or two children out. Uniform crossover picks each section independently from one parent or the other; single-point crossover takes a prefix of sections from one parent and the suffix from the other. The bet crossover makes is that fitness is at least partly **separable** — that the contribution of a good constraints block is largely independent of which exemplars sit beneath it. When that bet holds, recombination is a genuine shortcut. When the prompt's parts interact strongly (an output format that only works because a specific exemplar demonstrates it), crossover produces incoherent children and mutation carries most of the progress. ## A worked setting Consider optimizing the prompt for a fraud-alert triage assistant that must label each alert as escalate, monitor or dismiss, with a one-line justification. After a few generations the population holds one parent whose constraint block is unusually good at catching the rare escalate cases, and another whose output-format block reliably produces the exact three-field structure downstream systems parse. Neither is the best overall prompt. A crossover child that takes the first parent's constraints and the second parent's format block can beat both immediately; a hill-climber holding only the better-scoring parent would have to rediscover the other's format improvement by chance. ## Selection and diversity Selection decides who reproduces. Truncation selection keeps the top-N by score; tournament selection samples a few candidates and takes the best of that sample, which gives weaker prompts a survival chance and preserves diversity. Elitism copies the single best prompt into the next generation unchanged so a lucky mutation cannot lose it. The dominant failure is **premature convergence**: strong selection pressure makes one lineage dominate within two or three generations, the population becomes near-clones, crossover between clones produces the parent again, and only mutation still generates novelty — an expensive way to run hill-climbing. Countermeasures are milder selection pressure, higher mutation rate, explicitly injecting fresh random candidates each generation, or measuring population diversity (pairwise dissimilarity of prompts) and reacting when it collapses. ## Cost Population methods are expensive: population size times evaluation items times generations, all in model calls. They are also naturally parallel — a whole generation can be scored concurrently — which makes them fit rate-limited APIs better than a strictly sequential chain of dependent steps. Against noisy scores, selection amplifies luck: a candidate that scored high on one lucky evaluation subset becomes a parent and spreads. Scoring all candidates in a generation on the *same* evaluation items makes the comparison paired and much less noisy. ## When to prefer this family Reach for evolutionary search when the prompt has several separable components you expect to improve independently, when the landscape looks rugged enough that a single climber keeps stalling, and when you have parallel budget to spend. Prefer a simpler beam or hill-climb when the prompt is short, the budget is small, or improvements are obviously sequential refinements of one idea.