Module 7: Prompt Engineering for Architecture

Lesson 5: Prompt Reuse and Versioning

In a "Certified" environment, you cannot have "Floating Prompts" (prompts hardcoded as strings in your application code). Why? Because when Anthropic releases a new model version (e.g., Sonnet 3.5 to 3.6), your old prompt might behave differently. You need a way to Track, Test, and Version your instructions.

In this lesson, we learn how to treat "Prompts as Code."

---

1. Prompt Templating (The Dry Principle)

"Don't Repeat Yourself." If you have 50 agents using the same "Role" (Module 7, Lesson 1), you should not copy-paste that role 50 times.

• The Solution: Use external files (e.g., .yaml or .json) to store the "Base" prompt and use placeholders for the dynamic data.
• Example: You are a {{ ROLE }}. Your goal is to {{ TASK }}.

---

2. Versioning Prompts with Git

Your prompts should live in your Git repository. This allows you to:

• Audit Changes: "Who changed the security guardrail last Tuesday?"
• Rollback: "The new prompt is causing hallucinations; let's revert to v2."
• Branching: Test a "Greedy" prompt on one branch while keeping the "Conservative" prompt on production.

---

3. The "Prompt Registry" Pattern

For enterprise-scale systems, architects build a Prompt Registry.

1. The code calls get_prompt("coder_agent", version="1.2").
2. The registry fetches the correctly formatted string from a central store.
3. This decouples the Logic of your app from the Language of your instructions.

---

4. Testing for "Model Drift"

When a model is updated, the same prompt can produce different results.

• The Architect's Fix: Create a "Baseline Test Suite." For every critical prompt, have a set of "Gold Standard" inputs and outputs. If the new model version changes the output by > 10%, your prompt needs a version update.

---

5. Summary of Module 7

Module 7 has mastered the "Mind" of the system.

• You used Role-based Prompting to set the standard (Lesson 1).
• You used Clarity and Specificity to eliminate ambiguity (Lesson 2).
• You used Decomposition to handle complex logic (Lesson 3).
• You used Guardrails to secure the output (Lesson 4).
• You used Versioning to make the system maintainable (Lesson 5).

In Module 8, we move to the next level of precision: Structured Output and Schema Design.

---

Interactive Quiz

1. Why should you avoid hardcoding prompt strings in your application code?
2. What is a "Gold Standard" test suite for prompts?
3. How does a Prompt Registry help with multi-model deployments (e.g., switching between Sonnet and Haiku)?
4. Create a YAML-based prompt template for a "Legal Auditor" agent. Include placeholders for contract_text and law_jurisdiction.

---

Reference Video: