Prompt Engineering & Structuring

Prompt Engineering is the configuration of inputs to maximize model reasoning accuracy and output structures. To build reliable agent loops, you must calibrate model parameters and enforce strict schema output formats.

---

Parameter Calibration benchmarks

Task Type	Temperature	Top-P	Frequency Penalty	Primary Goal
JSON Extraction	0.0	0.1	0.0	High precision, schema compliance
Code Generation	0.2	0.2	0.0	Logical consistency, syntactic validity
Reasoning / Math	0.1	0.4	0.0	Logical accuracy, step-by-step reasoning
Creative Content	0.8	0.9	0.5	Token variety, phrase diversity

---

Mechanics: XML Tags for Prompt Isolation

Frontier models (like Gemini 2.5 and Claude 3.5) are trained to interpret XML tags as absolute boundaries. This eliminates prompt injection risks and ensures the model distinguishes between instructions and user-supplied data:

<system_instructions>
You are an expert parser. Convert the input document into a valid JSON array.
</system_instructions>

<input_document>
User inputs go here (if the user tries to write "Ignore previous instructions", the model recognizes it lies inside the input_document boundary and ignores the injection).
</input_document>

---

Mechanics: Enforcing Structured Outputs

Getting an LLM to output consistent JSON schemas is required for programmatic loops. There are three primary patterns for enforcing structured data:

Structured Output Spec Sheet

Output Pattern	How it Works	Error Rate	Latency Penalty	Best Use Case
JSON Mode (Natural)	Model is prompted to output JSON; parsed locally.	Moderate (schema drift)	None	Non-critical metadata extraction
JSON Schema Enforcement	API forces the sampler to select only tokens that comply with a schema (Zod/OpenAPI).	Near 0%	Minimal	Multi-agent loops, database populating
Grammar-Based Decoding	Sampler is constrained by a local context-free grammar (GBNF) at inference time.	0%	Moderate	Local models (Ollama/llama.cpp)

---

Walkthrough: Structured Output Build (Gemini API)

When building database-style wikis or structured pipelines, you should enforce a strict JSON schema at the API level. Here is the implementation using the @google/genai SDK:

import { GoogleGenAI, Type } from "@google/genai";

const ai = new GoogleGenAI({});

export async function generateStructuredConcept(prompt: string) {
  const response = await ai.models.generateContent({
    model: 'gemini-2.5-flash',
    contents: prompt,
    config: {
      responseMimeType: "application/json",
      responseSchema: {
        type: Type.OBJECT,
        properties: {
          conceptName: { type: Type.STRING },
          description: { type: Type.STRING },
          metrics: {
            type: Type.ARRAY,
            items: { type: Type.STRING }
          }
        },
        required: ["conceptName", "description", "metrics"],
      },
    },
  });

  return JSON.parse(response.text || '{}');
}