Prompt Optimizer

A structured framework and toolset for engineering, optimizing, and evaluating LLM system and user prompts to achieve higher determinism, fewer hallucinations, and lower token usage.

Core Capabilities

1. User Prompt Optimization (optimize-user-prompt):
   • Converts ambiguous user requests into clear, structured, contextual instructions.
   • Decomposes complex multi-part tasks into step-by-step reasoning sequences.
   • Specifies explicit negative constraints (what NOT to do) to eliminate hallucinations.
2. System Prompt Engineering (optimize-system-prompt):
   • Defines strict persona, role, tone, and operational boundaries.
   • Embeds few-shot input/output examples with realistic edge cases.
   • Enforces deterministic output formats (JSON schemas, Markdown structures).
   • Optimizes instruction density to prevent model attention degradation.
3. Iterative Prompt Refinement (iterate-prompt):
   • Tests candidate prompts against edge-case inputs.
   • Evaluates outputs against rubric criteria: correctness, conciseness, instruction following, and safety.
   • Performs ablation: trims redundant verbiage while preserving functional semantics.

Systematic Prompt Optimization Workflow

[Input Prompt] ──> [Task Clarification] ──> [Format & Constraint Enforcement]
      │                                                │
      ▼                                                ▼
[Final Prompt] <── [Token Trimming] <── [Few-Shot Exemplar Synthesis]

Step 1: Clarify Intent & Objectives

• Identify primary goal, target audience/model, and expected input variables.
• Eliminate vague adjectives ("good", "detailed") in favor of specific criteria ("includes time complexity, 3 test cases").

Step 2: Establish Constraints & Guardrails

• Positive Instructions: Define exactly what steps the model must execute.
• Negative Constraints: Explicitly prohibit common model tendencies (e.g., "Do not apologize", "Do not include placeholder comments").
• Error Handling: Instruct the model on how to handle missing data or ambiguous queries.

Step 3: Few-Shot Exemplar Engineering

• Include 2-3 high-quality input/output pairs demonstrating desired output formatting and reasoning depth.
• Ensure at least one example covers a non-standard or edge-case input.

Step 4: Token Optimization

• Compress repetitive phrasing without changing instruction meaning.
• Use bullet points and hierarchical markdown headers for high attention retention.