Playbook/Stage 04

Go Deeper

Level up the craft

Prompt Engineering

The craft of getting reliable, high-quality outputs from AI models.

Builder’s checkYour prompts are more of your product than you think. If a competitor cloned your interface tomorrow but not your system prompts, would they have your product, or just your paint job? The prompt IS the moat, and most builders treat it as an afterthought.

Prompts are the interface between what you want and what the model does. A well-crafted prompt can turn a cheap model into a great product. A lazy prompt will waste the most expensive model on earth.

The anatomy of a prompt

ComponentWhat it doesExample
System promptSets persona, constraints, behavioral rules“You are a senior tax advisor. Never give advice without citing the relevant tax code.”
ContextBackground information the model needsThe user's financial data, relevant regulations
InstructionThe actual task“Analyze this return and identify the three highest-risk deductions.”
ExamplesDemonstrations of desired input/outputTwo or three sample analyses
Output formatHow to structure the response“Respond in JSON with fields: deduction, risk_level, explanation.”

Three strategies that cover 90% of use cases

Zero-shot prompting. Give the model an instruction with no examples. Works for simple, well-defined tasks. The advantage is simplicity. The disadvantage is inconsistent formatting.

Few-shot prompting. Provide 2-5 examples before your actual query. Use this when you need specific output formats or domain-specific terminology. Three to five examples is usually enough.

Chain of Thought (CoT). Ask the model to reason step by step before giving its final answer. This dramatically improves performance on math, multi-step reasoning, and complex analysis.

As a rough guide: zero-shot gets you about 60% reliability, few-shot pushes that to 80%, and chain of thought reaches 90% or higher.

System prompts: where your product lives

The system prompt is the single most important piece of text in your entire application. An effective system prompt covers four things:

  1. Identity. Who or what is the assistant?
  2. Constraints. What should the model NOT do?
  3. Behavior. How should it interact?
  4. Format. How should responses be structured?

Version control your prompts. Treat them like code. Track which version produced which outputs. A “small tweak” to a system prompt can change behavior in ways you don't expect.

Getting structured output

For production systems, you almost always need the model to return data in a specific format. Three approaches: JSON mode (most APIs support forcing valid JSON), Function calling / tool use (define schemas the model must follow), and Structured output libraries (like Instructor for Python).

Prompt injection: the security risk you cannot ignore

When your prompts include user input, you're at risk of prompt injection. There is no perfect defense, but effective mitigations include: delimiters, instruction hierarchy, output filtering, and least privilege.

Test your prompts systematically. Build an eval set covering the happy path, edge cases, and adversarial inputs. Track accuracy, consistency, latency, and token usage.

/system-prompt

System Prompt Architect

Writes and pressure-tests your AI product's system prompt.

skill
---
description: Design, write, and pressure-test a system prompt for your AI product feature.
---

You are a system prompt architect. The user is building an AI product feature and needs a production-grade system prompt.

Ask: "What does this AI feature do, and who uses it? Be specific."

Then build through four layers: 1. Identity, 2. Constraints, 3. Behavior, 4. Output format.

After writing the prompt, run three stress tests:
1. An adversarial input
2. An edge case
3. A request outside scope

End with: "Version control this prompt. Treat changes like code changes."

Reference: https://builderspath.dev/playbook/#prompt-engineering