Automatic Prompt & Context Engineering

"Why write prompts by hand when you can write a program that writes better prompts than you ever could?"

Prompt, Meta-Prompting AI Agent

12.5.1 From Manual Craft to Programmatic Optimization

Manual prompt engineering hits a ceiling. Prompts tuned for GPT-4o regularly underperform on Claude or Gemini. Prompts optimized for one set of examples often fail to generalize to new data. As systems grow more complex, maintaining dozens of hand-tuned prompts turns into a real engineering burden, not a hypothetical one.

Automatic prompt optimization addresses these issues by treating prompt design as a search problem. The core idea is simple: define a task metric (accuracy, F1 score, user satisfaction), provide a set of training examples with expected outputs, and use an optimization algorithm to search over the space of possible prompts to find one that maximizes the metric. The key insight is that the search space is text (not continuous parameters), so the optimization algorithms must work with discrete, natural language modifications.

Figure 12.5.1: The DSPy optimization workflow. You define what you want (signatures, modules, metrics), the optimizer searches over prompt variations using training examples, and the compiler produces an optimized prompt. This inverts the traditional ML workflow: instead of optimizing model weights, you optimize the prompt text.

12.5.2 DSPy: Declarative Prompting with Optimizers

DSPy (Khattab et al., 2024) is the most comprehensive framework for automatic prompt optimization. It introduces a programming model where you declare what you want the LLM to do (as a signature specifying inputs and outputs) and let an optimizer figure out how to prompt the model to achieve it. This separates the task specification from the prompt implementation, making systems portable across models.

A DSPy program consists of three components: Signatures define the input-output contract (e.g., "question, context to answer"), Modules implement reasoning patterns (ChainOfThought, ReAct, ProgramOfThought), and Optimizers (called "teleprompters") tune the prompts and few-shot examples to maximize a metric. The optimizer evaluates the program on training examples, identifies which components need improvement, and iteratively refines the prompt instructions and selected demonstrations.

# DSPy: Declarative prompt optimization
import dspy
# Configure the language model
lm = dspy.LM("openai/gpt-4o-mini")
dspy.configure(lm=lm)
# Define a signature: inputs and outputs
class AnswerQuestion(dspy.Signature):
    """Answer a factual question using the provided context."""
    context: str = dspy.InputField(desc="relevant passages from the knowledge base")
    question: str = dspy.InputField(desc="the user's question")
    answer: str = dspy.OutputField(desc="a concise, factual answer")
    # Create a module that uses chain-of-thought reasoning
    cot_qa = dspy.ChainOfThought(AnswerQuestion)
    # Use the module (before optimization)
    result = cot_qa(
        context="The Eiffel Tower was completed in 1889 for the World's Fair.",
        question="When was the Eiffel Tower built?"
    )
    print(f"Answer: {result.answer}")
# Now optimize with MIPROv2
from dspy.teleprompt import MIPROv2
# Define a metric function
def answer_accuracy(example, prediction, trace=None):
    """Check if the predicted answer matches the expected answer."""
    return prediction.answer.lower().strip() == example.answer.lower().strip()
# Provide training examples
trainset = [
    dspy.Example(
    context="Python was created by Guido van Rossum in 1991.",
    question="Who created Python?",
    answer="Guido van Rossum"
    ).with_inputs("context", "question"),
    dspy.Example(
    context="The speed of light is approximately 299,792,458 m/s.",
    question="What is the speed of light?",
    answer="approximately 299,792,458 m/s"
    ).with_inputs("context", "question"),
    # ... more training examples
]
# Run the optimizer
optimizer = MIPROv2(metric=answer_accuracy, num_threads=4)
optimized_qa = optimizer.compile(cot_qa, trainset=trainset)
# The optimized module has tuned instructions and selected demonstrations
print("Optimized prompt instructions:")
print(optimized_qa.predict.signature.instructions)

DSPy's key optimizers include: BootstrapFewShot, which selects and orders the best few-shot demonstrations from a training set; MIPROv2 (Multi-prompt Instruction Proposal Optimizer), which jointly optimizes instructions and demonstrations using Bayesian search; and BootstrapFewShotWithRandomSearch, which combines demonstration selection with random perturbations to the prompt structure. MIPROv2 is generally the recommended starting point because it optimizes both the instruction text and the example selection simultaneously.

12.5.3 OPRO: LLM-Based Prompt Optimization

OPRO (Optimization by PROmpting, Yang et al. 2024) takes a conceptually elegant approach: use an LLM to optimize prompts for another LLM (or the same LLM). The optimizer LLM receives a "meta-prompt" containing the task description, previously tried prompts, and their scores on an evaluation set. It then generates new candidate prompts, which are evaluated on the task, and the results feed back into the meta-prompt. Over multiple iterations, the optimizer converges on high-performing prompts.

The OPRO workflow is straightforward: (1) start with an initial prompt (which can be empty or a human-written draft), (2) evaluate the prompt on a validation set, (3) present the optimizer LLM with the prompt history and scores, (4) the optimizer generates new candidate prompts, (5) evaluate the candidates and add them to the history, (6) repeat until convergence or a budget limit is reached. The meta-prompt instructs the optimizer to analyze patterns in successful versus unsuccessful prompts and generate improvements.

# OPRO-style prompt optimization (simplified implementation)
from openai import OpenAI
import json
client = OpenAI()
def evaluate_prompt(prompt: str, eval_set: list) -> float:
    """Evaluate a prompt on the evaluation set, returning accuracy."""
    correct = 0
    for example in eval_set:
        response = client.chat.completions.create(
            model="gpt-4o-mini",
            messages=[
            {"role": "system", "content": prompt},
            {"role": "user", "content": example["input"]},
            ],
            )
        predicted = response.choices[0].message.content.strip()
        if example["expected"].lower() in predicted.lower():
            correct += 1
            return correct / len(eval_set)
        def opro_optimize(task_desc: str, eval_set: list, n_iterations: int = 10) -> str:
            """Use an LLM to iteratively optimize a prompt."""
            history = []
            # Start with a basic prompt
            current_prompt = f"You are a helpful assistant. {task_desc}"
            score = evaluate_prompt(current_prompt, eval_set)
            history.append({"prompt": current_prompt, "score": score})
            for i in range(n_iterations):
                # Build the meta-prompt with history
                history_text = "\n".join([
                    f"Prompt: {h['prompt']}\nScore: {h['score']:.2f}"
                    for h in sorted(history, key=lambda x: x["score"])[-10:]
                    ])
                # Ask the optimizer LLM to generate a better prompt
                response = client.chat.completions.create(
                    model="gpt-4o",
                    messages=[
                    {"role": "system", "content": """You are a prompt optimization expert.
                    Analyze the history of prompts and their scores below. Generate a new prompt
                    that will score higher. Focus on what patterns made high-scoring prompts
                    succeed and low-scoring prompts fail. Return ONLY the new prompt text."""},
                    {"role": "user", "content": f"Task: {task_desc}\n\n"
                    f"Previous attempts (sorted by score):\n{history_text}\n\n"
                    f"Generate an improved prompt:"},
                    ],
                    )
                new_prompt = response.choices[0].message.content.strip()
                new_score = evaluate_prompt(new_prompt, eval_set)
                history.append({"prompt": new_prompt, "score": new_score})
                print(f"Iteration {i+1}: score={new_score:.2f}")
                # Return the best prompt
                best = max(history, key=lambda x: x["score"])
                return best["prompt"]
                # Example usage
                best_prompt = opro_optimize(
                    task_desc="Classify customer emails as positive, negative, or neutral sentiment.",
                    eval_set=[
                    {"input": "Great product, love it!", "expected": "positive"},
                    {"input": "Terrible service, never again.", "expected": "negative"},
                    {"input": "Order arrived on Tuesday.", "expected": "neutral"},
                    # ... more examples
                    ],
                    n_iterations=5
                    )

12.5.4 TextGrad and Gradient-Based Text Optimization

TextGrad (Yuksekgonul et al., 2024) brings the concept of backpropagation to text optimization. In numerical deep learning, gradients indicate how to adjust parameters to reduce loss. TextGrad computes "textual gradients": natural language feedback that describes how to modify a text to improve a given objective. The LLM serves as both the function being optimized and the source of gradient information.

The workflow is analogous to numerical optimization: (1) a forward pass generates an output from the current prompt, (2) a loss function evaluates the output quality, (3) a backward pass generates textual feedback describing how to improve the prompt, and (4) an update step applies the feedback to produce a new prompt version. This process repeats until the output quality converges.

TextGrad's advantage is generality: it can optimize any text variable in a compound system, not just the prompt instruction. It can simultaneously tune the system prompt, few-shot examples, output format specification, and even the retrieval query in a RAG pipeline. This makes it suitable for optimizing complex, multi-component LLM systems where multiple text variables interact.

12.5.5 EvoPrompt: Evolutionary Prompt Search

EvoPrompt (Guo et al., 2024) applies evolutionary algorithms to prompt optimization. Starting with a population of prompt candidates (which can be human-written or randomly generated), the algorithm uses selection, crossover, and mutation operations to evolve prompts over multiple generations. Selection keeps the highest-scoring prompts; crossover combines elements from two successful prompts; and mutation introduces random variations through LLM-powered paraphrasing.

The evolutionary approach has several advantages. It maintains diversity in the prompt population, reducing the risk of converging on a local optimum. It can discover non-obvious prompt formulations that neither humans nor gradient-based methods would produce. And it is embarrassingly parallel: each candidate in the population can be evaluated independently, making it efficient to run on modern hardware.

12.5.6 Prompt Compression: LLMLingua and LongLLMLingua

As context windows grow and RAG systems retrieve increasing amounts of context, managing what fits within the model's attention becomes critical. Prompt compression reduces the token count of prompts and retrieved contexts while preserving the information needed for accurate responses.

LLMLingua (Jiang et al., 2023) uses a small language model (such as GPT-2 or LLaMA-7B) to identify and remove tokens that contribute least to the prompt's information content. The approach computes the perplexity contribution of each token and removes those with the lowest information density. This can achieve 2x to 20x compression ratios with minimal quality loss, depending on the redundancy of the original text.

LongLLMLingua extends this approach specifically for RAG scenarios where long retrieved contexts need compression. It adds document-level relevance estimation (prioritizing passages most relevant to the query), a "contrastive perplexity" metric that preserves query-relevant tokens, and a dynamic compression ratio that applies more aggressive compression to less relevant passages. This makes it particularly effective for reducing the cost and latency of RAG systems while maintaining answer quality.

# Prompt compression with LLMLingua
# Install: pip install llmlingua
from llmlingua import PromptCompressor
# Initialize the compressor with a small model
compressor = PromptCompressor(
    model_name="microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank",
    use_llmlingua2=True,
)
# Example: compress a long RAG context
original_context = """
The transformer architecture was introduced in the seminal paper 'Attention Is All
You Need' by Vaswani et al. in 2017. The key innovation was the self-attention
mechanism, which allows the model to weigh the importance of different parts of
the input sequence when producing each output element. This replaced the recurrent
connections used in previous sequence-to-sequence models like LSTMs and GRUs.
The original transformer used an encoder-decoder architecture with multi-head
attention layers. The encoder processes the input sequence in parallel (unlike
RNNs which process sequentially), and the decoder generates the output sequence
one token at a time, attending to both the encoder output and previously generated
tokens. Position information is injected through sinusoidal positional encodings.
Subsequent work showed that using only the decoder (GPT series) or only the encoder
(BERT) could achieve strong results on generation and understanding tasks respectively.
The scaling laws discovered by Kaplan et al. (2020) demonstrated that model performance
improves predictably with increases in model size, dataset size, and compute budget.
"""
question = "What is the key innovation of the transformer architecture?"
# Compress the context
compressed = compressor.compress_prompt(
    context=[original_context],
    instruction="",
    question=question,
    rate=0.5, # target 50% compression
)
print(f"Original tokens: {compressed['origin_tokens']}")
print(f"Compressed tokens: {compressed['compressed_tokens']}")
print(f"Compression ratio: {compressed['ratio']:.1f}x")
print(f"Compressed text: {compressed['compressed_prompt']}")

12.5.7 Context Engineering: MCP and Dynamic Context Assembly

Context engineering is the practice of dynamically assembling the right information to include in each LLM call. While prompt engineering focuses on how to instruct the model, context engineering focuses on what information the model needs to see. This is especially important for agent systems and complex applications where the relevant context changes with every interaction.

The Model Context Protocol (MCP), introduced by Anthropic, standardizes how applications provide context to language models. MCP defines a protocol through which tools, data sources, and context providers expose information to the model in a structured way. Rather than hardcoding context retrieval into the prompt template, MCP allows the model to request specific context from registered providers at runtime. This creates a clean separation between the model's reasoning and the context supply chain. MCP is covered in detail in Chapter 27.

Dynamic context assembly involves selecting, filtering, and ordering context at inference time based on the current query. A well-designed context assembly pipeline: (1) retrieves candidate context from multiple sources (vector stores, knowledge graphs, APIs, conversation history), (2) ranks candidates by relevance to the current query, (3) applies compression to fit within the context window budget, (4) orders the context to place the most important information in positions where the model attends most strongly (typically the beginning and end of the context), and (5) adds structural markers (headers, separators) that help the model parse the assembled context.

12.5.8 When to Use Automatic vs. Manual Prompt Engineering

Automatic prompt optimization is not always the right choice. The decision depends on the task complexity, the number of use cases, and the available evaluation data.

Use manual prompt engineering when: you are prototyping a new application and the task definition is still evolving; you have fewer than 20 evaluation examples; the task requires nuanced judgment that is difficult to capture in a metric; or you need full control over the prompt for regulatory or compliance reasons (some regulated industries require that all prompts be human-reviewed and approved).

Use automatic prompt optimization when: you have a stable task definition with a clear metric; you have at least 50 to 100 labeled examples for optimization (and a held-out test set for evaluation); you need to support multiple models or model versions; you are maintaining many prompts and manual tuning does not scale; or you observe that small prompt variations cause large performance swings (indicating that the current prompt is not robust).

Use prompt compression when: your RAG system retrieves more context than fits in the model's context window; you are paying per token and context is a significant cost driver; or you observe that longer contexts are degrading answer quality (the "lost in the middle" problem, where models struggle to use information in the middle of long contexts).

12.5.9 Lab: Beat Your Hand-Written Prompt with DSPy

Exercises

from llmlingua import PromptCompressor
compressor = PromptCompressor("microsoft/llmlingua-2-xlm-roberta-large-meetingbank", device_map="cpu")
out = compressor.compress_prompt(retrieved_docs, rate=0.5)
compressed_context = out["compressed_prompt"]
answer = llm.chat(template.format(context=compressed_context, q=user_q))