Autonomous Prompt Optimization

Prompt includes a suite of autonomous engines that improve prompts without manual intervention. These systems cover the full optimization lifecycle — from single-pass rule-based fixes to multi-generation genetic evolution.

Overview

Engine	Purpose	Approach
`PromptAutoImprover`	One-shot prompt rewriting	Multi-pass rule analysis across 9 categories
`PromptAutopilot`	Iterative autonomous refinement	Diagnose → fix → score loops with strategy selection
`PromptEvolutionEngine`	Genetic algorithm optimization	Population-based evolution with crossover and mutation
`PromptFeedbackLoop`	Feedback-driven improvement	Collects output feedback, detects patterns, suggests fixes

PromptAutoImprover

Takes a raw prompt and produces concrete improved versions through multiple analysis passes. Each pass identifies issues in a specific category, rewrites the prompt, and explains what changed.

Quick Start

using Prompt;

var improver = new PromptAutoImprover();
var result = improver.Improve("Tell me about dogs", new ImprovementConfig
{
    Intensity = ImprovementIntensity.Moderate,
    MaxTokenBudget = 500
});

Console.WriteLine($"Before: {result.OriginalScore:F2}");
Console.WriteLine($"After:  {result.ImprovedScore:F2}");
Console.WriteLine($"Improved prompt:\n{result.ImprovedText}");

foreach (var pass in result.Passes)
{
    Console.WriteLine($"  [{pass.Category}] {pass.Explanation}");
}

Improvement Categories

Category	What it fixes
`Specificity`	Replaces vague language with precise instructions
`Clarity`	Simplifies wording for better model comprehension
`Structure`	Adds missing role, format, and constraint elements
`TokenEfficiency`	Reduces token count without losing meaning
`Completeness`	Adds missing instructions the model likely needs
`AntiPattern`	Removes known patterns that degrade output quality
`FormatSpec`	Improves output format specification
`ExampleQuality`	Adds or improves few-shot examples
`SafetyGuardrail`	Strengthens guardrails and safety constraints

Intensity Levels

Light — Conservative fixes for clear issues only
Moderate — Balanced improvements across categories (default)
Deep — Aggressive rewriting for maximum quality gain

Configuration

var config = new ImprovementConfig
{
    Intensity = ImprovementIntensity.Deep,
    MaxTokenBudget = 1000,           // Upper bound on improved prompt length
    FocusCategories = new[]          // Limit to specific categories (optional)
    {
        ImprovementCategory.Specificity,
        ImprovementCategory.Structure
    }
};

HTML Reports

The engine can generate a self-contained HTML report showing before/after scores, per-pass changes, and quality breakdowns:

string html = result.ToHtml();
File.WriteAllText("improvement-report.html", html);

PromptAutopilot

An iterative refinement engine that runs a diagnose → fix → score loop until the prompt meets a target quality bar or the generation budget is exhausted.

Quick Start

using Prompt;

var autopilot = new PromptAutopilot();
var result = autopilot.Run("Summarize this article", new AutopilotConfig
{
    Strategy = AutopilotStrategy.WorstFirst,
    MaxGenerations = 10,
    TargetScore = 0.85
});

Console.WriteLine($"Generations: {result.Generations}");
Console.WriteLine($"Final score: {result.FinalScore:F2}");
Console.WriteLine($"Final prompt:\n{result.FinalText}");

Strategies

Strategy	Behavior
`WorstFirst`	Fix the most severe issue first each generation
`Shotgun`	Fix all issues in a single pass
`Rotating`	Focus on one quality dimension per generation (cycling through)
`Conservative`	Apply only high-confidence, low-risk fixes
`Aggressive`	Aggressive rewriting targeting maximum score gain

Key Classes

AutopilotDiagnosis — A single finding from analysis (category, severity, description, suggested fix)
AutopilotGeneration — Snapshot of one refinement iteration (prompt text, score, fixes applied)
AutopilotResult — Full history of the refinement run with before/after comparison

PromptEvolutionEngine

A genetic-algorithm optimizer that evolves a population of prompt variants over multiple generations. The engine applies crossover, mutation, and selection to discover high-fitness prompt formulations.

Quick Start

using Prompt;

// Seed population with variations of your prompt
var seeds = new[]
{
    "Explain quantum computing in simple terms.",
    "Describe quantum computing for a beginner.",
    "What is quantum computing? Keep it simple."
};

var engine = new PromptEvolutionEngine();
var result = engine.Evolve(seeds, new EvolutionConfig
{
    PopulationSize = 20,
    MaxGenerations = 50,
    MutationRate = 0.15,
    CrossoverRate = 0.7,
    EliteCount = 2,
    FitnessTargetThreshold = 0.95
});

Console.WriteLine($"Best: {result.BestOrganism.Text}");
Console.WriteLine($"Fitness: {result.BestOrganism.Fitness:F3}");
Console.WriteLine($"Generations: {result.TotalGenerations}");

Mutation Strategies

Strategy	Effect
`WordShuffle`	Randomly reorder words within a sentence
`SynonymReplace`	Replace common words with synonyms
`SentenceReorder`	Reorder sentences in the prompt
`ToneShift`	Shift formality level
`Compress`	Remove filler words and redundancy
`Expand`	Add clarifying or emphatic phrases
`InstructionRephrase`	Rephrase imperative sentences with different verb forms

Crossover Methods

Method	Behavior
`SinglePoint`	Split at midpoint sentence boundary
`Uniform`	Randomly pick sentences from each parent
`SemanticBlend`	Interleave sentences alternating parents

Selection Methods

Method	Behavior
`Tournament`	Tournament selection among random candidates (default)
`RouletteWheel`	Probability proportional to fitness
`RankBased`	Rank-based probability with linear ranking
`Elitism`	Top-N organisms pass through unchanged

Configuration

var config = new EvolutionConfig
{
    PopulationSize = 30,          // Organisms per generation (min 4)
    MaxGenerations = 100,         // Max generations to run
    MutationRate = 0.2,           // Base mutation probability per offspring
    CrossoverRate = 0.8,          // Crossover probability per pair
    EliteCount = 3,               // Top organisms that pass unchanged
    Selection = SelectionMethod.Tournament,
    TournamentSize = 4,           // Candidates per tournament
    AdaptiveMutationRate = true,  // Auto-adjust mutation to diversity
    FitnessTargetThreshold = 0.95 // Early-stop fitness target
};

Key Classes

PromptOrganism — An individual prompt variant with text, fitness score, generation, parentage, and mutation history
GenerationStats — Per-generation statistics (best/avg/worst fitness, diversity, stagnation count)
EvolutionResult — Full evolution run output with best organism, history, population, and summary

PromptFeedbackLoop

Collects structured feedback on prompt outputs, detects recurring issue patterns through keyword frequency and category analysis, and autonomously suggests (or auto-applies) prompt refinements.

Quick Start

using Prompt;

var loop = new PromptFeedbackLoop();

// Record feedback on prompt outputs
loop.AddFeedback(new PromptFeedback
{
    PromptId = "summarizer-v1",
    Rating = FeedbackRating.Poor,
    Category = "accuracy",
    Comment = "The summary missed key financial figures"
});

loop.AddFeedback(new PromptFeedback
{
    PromptId = "summarizer-v1",
    Rating = FeedbackRating.Acceptable,
    Category = "accuracy",
    Comment = "Numbers were approximate but acceptable"
});

// Get analysis and suggestions
var analysis = loop.Analyze("summarizer-v1");
Console.WriteLine($"Health score: {analysis.HealthScore:F2}");

foreach (var suggestion in analysis.Suggestions)
{
    Console.WriteLine($"  [{suggestion.Category}] {suggestion.Description}");
}

Feedback Ratings

Rating	Value	Meaning
`Excellent`	5	Outstanding output
`Good`	4	Good with minor issues
`Acceptable`	3	Meets minimum bar
`Poor`	2	Below expectations
`Terrible`	1	Unacceptable output

Features

Pattern detection — Identifies recurring issues through keyword frequency and category analysis
Health scoring — Aggregate quality score based on feedback distribution
Timeline analysis — Tracks how prompt quality trends over time
Category heatmaps — Visualize which issue categories are most common
Proactive insights — Automatically surfaces emerging problems before they become patterns
Multi-format export — Export analysis as JSON, HTML, or Markdown

Combining the Engines

These tools work well together in a prompt optimization pipeline:

// 1. Start with AutoImprover for quick structural fixes
var improver = new PromptAutoImprover();
var improved = improver.Improve(rawPrompt, new ImprovementConfig
{
    Intensity = ImprovementIntensity.Moderate
});

// 2. Feed improved versions into the Evolution Engine for exploration
var engine = new PromptEvolutionEngine();
var evolved = engine.Evolve(
    new[] { improved.ImprovedText, rawPrompt },
    new EvolutionConfig { MaxGenerations = 30 }
);

// 3. Use Autopilot for final polishing
var autopilot = new PromptAutopilot();
var final = autopilot.Run(evolved.BestOrganism.Text, new AutopilotConfig
{
    Strategy = AutopilotStrategy.Conservative,
    TargetScore = 0.9
});

// 4. Deploy and collect feedback
var loop = new PromptFeedbackLoop();
// ... record production feedback, analyze, iterate

Table of Contents

Autonomous Prompt Optimization

Overview

PromptAutoImprover

Quick Start

Improvement Categories

Intensity Levels

Configuration

HTML Reports

PromptAutopilot

Quick Start

Strategies

Key Classes

PromptEvolutionEngine

Quick Start

Mutation Strategies

Crossover Methods

Selection Methods

Configuration

Key Classes

PromptFeedbackLoop

Quick Start

Feedback Ratings

Features

Combining the Engines

See Also