Don't Panic. Just Program.

The practical guide to Stanford's DSPy framework. Stop writing prompts. Start writing code.

Launch price — goes to $34 after early access

7 chapters · 6,000+ lines of real code · Free lifetime updates

The old way

  • ✗Your prompt worked great in testing.
  • ✗Then you upgraded the model.
  • ✗Then your client changed the use case.
  • ✗Then it started hallucinating on edge cases.
  • ✗Then you added "IMPORTANT:" for the fourth time.
  • ✗Then you had no idea why it worked Monday and broke Tuesday.

The DSPy way

  • ✓You write a Python class.
  • ✓DSPy writes the prompt.
  • ✓You swap models with one line.
  • ✓You run an optimizer — prompts improve automatically.
  • ✓You write tests like normal code.
  • ✓You ship with confidence.

DSPy is a framework for programming LLMs, not prompting them.

When you prompt an LLM, you write instructions in English and hope the model interprets them the way you intended. When you program with DSPy, you declare what you want — inputs, outputs, types, constraints — and let the framework figure out how to ask for it. You write Python. DSPy writes the prompts. And when you need better prompts, you run an optimizer and DSPy rewrites them automatically based on examples that actually work.

It's the difference between manually tuning a SQL query every time your data changes versus writing a proper ORM. One is maintenance. The other is engineering.

# Before
prompt = f"""You are a
helpful AI assistant.
Analyze this startup pitch.
Respond in JSON format.
IMPORTANT: Only valid JSON.
SERIOUSLY: Just the JSON.

Pitch: {pitch}"""
# After
class AnalyzeStartup(dspy.Signature):
    """Analyze a startup pitch."""
    pitch: str = dspy.InputField()
    score: int = dspy.OutputField()
    strengths: list[str] = dspy.OutputField()
    weaknesses: list[str] = dspy.OutputField()
    verdict: str = dspy.OutputField()
Built on Stanford researchModel-agnostic — OpenAI, Anthropic, Gemini, local LLMsUsed in production at scale

See it happen.

Press Compile. Watch DSPy turn a Python class into a prompt.

YOUR CODE

class AnalyzeStartup(dspy.Signature):
  """Analyze a startup pitch."""

  pitch: str = dspy.InputField()
  viability_score: int = dspy.OutputField()
  strengths: list[str] = dspy.OutputField()
  weaknesses: list[str] = dspy.OutputField()
  verdict: str = dspy.OutputField()

DSPy ENGINE

Ready

THE PROMPT

This book is for you if...

🔧

You've been hacking with the OpenAI API

You can make an LLM do things. Your prompts work — until they don't. You've copy-pasted "respond only in JSON" into six different projects and you're starting to suspect there's a better way.

🚀

You're building something real

You need reliability, testability, and the ability to swap models without rewriting everything. You want LLM code that behaves like normal code — versioned, tested, debuggable.

😤

You're tired of prompt archaeology

You've spent hours tweaking wording, adding emphatic ALL-CAPS instructions, and still can't explain why a two-word change broke your pipeline. There's a better mental model — and this book teaches it.

What's Inside

Every chapter builds something real — not a toy, not a demo.

github.com/aliirz/harmless-dspy

All working code from every chapter — clone & run

1
Free
Ch. 1

Don't Panic

Your first working DSPy program.

  • ›Set up DSPy with any LLM provider
  • ›Write Signatures, Modules, and your first optimizer
  • ›Build a Startup Idea Roaster from scratch
2
Ch. 2

The Restaurant at the End of the Pipeline

Build a complete multi-step pipeline.

  • ›Compose modules into real pipelines
  • ›Add evaluation metrics that actually measure quality
  • ›Optimize prompts automatically with BootstrapFewShot
3
Ch. 3

Life, the Universe, and Retrieval

RAG done right — not the tutorial version.

  • ›Build a codebase Q&A system with DSPy + ChromaDB
  • ›Chunk, embed, and retrieve with optimizable modules
  • ›Evaluate retrieval quality with precision metrics
4
Ch. 4

The Babel Fish — Optimizers Demystified

Stop hand-writing prompts. Let DSPy find better ones automatically.

  • ›LabeledFewShot, BootstrapFewShot, MIPROv2, SIMBA — the full zoo
  • ›Build and evaluate a Ticket Classifier
  • ›The optimizer decision tree — know which to reach for and when
5
Ch. 5

So Long, and Thanks for All the Prompts

Agents that actually work in production.

  • ›Build tool-using agents with DSPy's ReAct module
  • ›Add guardrails so agents don't go rogue
  • ›Test and optimize agent behavior like regular code
6
Ch. 6

Mostly Harmless (in Production)

Ship it. Monitor it. Sleep at night.

  • ›Deploy behind FastAPI with caching and fallbacks
  • ›Track costs, latency, and quality per-request
  • ›Save and load optimized modules as artifacts
7
Ch. 7

The Answer Is 42 (Tokens)

Multimodal, custom optimizers, and the cutting edge.

  • ›Feed images to DSPy with dspy.Image
  • ›Build model cascades — cheap models for easy work, expensive for hard
  • ›Write custom teleprompters for your specific use case

About the Author

AR

Ali Raza

I built production DSPy systems before writing this book. I've hit every error message, debugged every silent failure, and rewritten every prompt that “worked in testing.” This book is the guide I wish I'd had.

Software engineer shipping AI tools at scale. Previously built developer platforms and open-source tools used by thousands.

github.com/aliirz·@aliirz on X

Stop writing prompts. Start writing code.

One-time purchase. Instant access. No subscription.

Buy once. Every future update is yours, free.

$24.99 launch price — goes to $34 after early access ends

✓7 full chapters — no filler✓6,000+ lines of real, runnable code✓Free lifetime updates — DSPy moves fast, the book keeps up✓Chapter 1 is free — no credit card needed

7-day money-back guarantee. Not what you expected? Email me within 7 days for a full refund, no questions asked.

Questions? ali@aliirz.com