The System Prompt That Controls an AI Coding Agent - Part 2, Section 3

📌 Video Description In this video (PART 2 – SECTION 3), we break down the design of a strict system prompt for an AI coding agent whose sole responsibility is to fix bugs — not explain them. We walk through how the prompt is constructed, why each rule exists, and how it enforces action-oriented, test-driven behavior from the agent. 🧠 What you’ll learn in this section: How an f-string is used to dynamically inject the working directory into a system prompt Defining a clear agent role: an AI coding agent that fixes bugs instead of describing them Enforcing a strict ENGLISH ONLY language constraint to avoid inconsistent outputs 🚨 Critical Rules Explained We analyze five critical rules that prevent common LLM failure modes, including: Forcing the agent to fix code using write_file() instead of explaining Ensuring entire files are rewritten, not partial snippets Preventing truncated or simplified code that can break production files Requiring the code to be fixed and tested successfully before responding Stopping execution immediately once the test passes These constraints ensure the agent behaves like a reliable automated developer, not a chatbot. 🔄 Workflow Enforcement The prompt defines a strict step-by-step workflow: List project files Read the full buggy file Rewrite the entire file with the fix applied Run tests and verify the expected output Stop immediately once the fix is confirmed This guarantees deterministic, repeatable behavior. 🐞 Bug Scenario Used The example bug is intentionally simple but realistic: A Celsius-to-Fahrenheit conversion function incorrectly multiplies by 2 The correct formula (C × 9/5 + 32) is enforced and verified via a test run The prompt also includes retry logic if the test fails, ensuring robustness. 🧩 Why this matters This approach is essential when building: Autonomous coding agents Tool-using LLM systems AI-driven refactoring or bug-fixing pipelines It demonstrates how prompt engineering + strict rules can turn an LLM into