Stop Playing Telephone with Your AI: A Structured Approach to Conversational Programming

Posted on | by ccadmin

Have you ever played telephone? A message passes from person to person until it reaches the last player, who compares what they heard to the original. The results are often hilarious, but in a company or organization where coworkers relay messages this way, the results could be costly and disastrous.

When you do conversational programming or vibecoding with an AI agent that writes your code, you’re playing telephone. This becomes especially difficult if you lack programming background, knowledge of language frameworks, or coding principles. Even experienced programmers who use vibecoding end up writing programs they can’t maintain or understand.

However, I believe programmers who have bad experiences with vibecoding are the same ones who don’t use best practices like test-driven development, agile, extreme programming, or DevOps. Organizations struggling with AI adoption are often the same ones struggling with Agile, Scrum, and Lean practices. It comes down to the telephone game — no contracts, no rules, no real structure for communicating safely.

Applying Engineering Discipline to Conversational Programming

In my experience with conversational programming (I prefer that term over vibecoding since that is what your doing, having a conversation with an AI), you must apply engineering discipline when having AI write code. Here are tips I find useful.

Start with a Well-Crafted Prompt

When developing an initial prompt, have a decent LLM write it. I first conceptualize what I want done, but understanding terminology and concepts are important. I recommend Vibe Coding: Building Production-Grade Software With GenAI, Chat, Agents, and Beyond by Gene Kim and Steve Yegge. They did studies on companies that successfully implemented vibecoding into their enterprise. What they found is they succeed because they use structured engineering approaches, applying DevOps principles—descendants of Extreme Programming using test-driven development, CI/CD pipelines, and testing tools.

Write a prompt, have an LLM agent refactor it, then record and archive it for future use. This creates a solid one-shot prompt.

Use AI Coding Agents, Not Chat Interfaces

Don’t use ChatGPT or chat-oriented interfaces to do a back and forth with a chat window and your IDE. Use AI coding agents like Windsurf, Cursor, Claude Code, or Cline. I personally use Claude Code with a subscription plan because I burn through many tokens, and Claude by Anthropic doesn’t place strict caps on token usage like API-based agents do.

Learn and Apply Test-Driven Development

Learn test-driven development concepts and include them in your prompts. TDD’s key tenet: write tests first. Know how programs or functions should behave and write tests around that.

TDD forces you to write programs in modular, testable ways. When your AI writes code, it tests for errors and rewrites until it works. For instance, without TDD, my Ionic app became a spaghetti mess—fixing one part broke another because dependencies and regressions weren’t tested. The blast radius of fixes affected other parts, growing to thousands of lines the code editor couldn’t contain.
In my github repo, I have a few applications that I have developed using TDD. I used AI coding agents to write the tests and then test the code against it.

Applying TDD to AI projects made code manageable and adding features easier. Modified modules had to pass tests, so the AI knew what broke and fixed it.

Use Configuration Files to Guide Your Coding Agent

Use various MD files to guide your agent. For instance, with Claude, there’s a CLAUDE.md file tuning agent behavior and an AGENT.md file with application instructions. Write separate MD files for architecture, coding, user interfaces, and so forth.

Leverage MCP (Model Context Protocol) Servers

MCP (Model Context Protocol) servers make AI coding agents more efficient. I spun up a Penpot server (a web-based graphic design tool), created an MCP server connecting to Penpot, had Claude Code connect to it, and using descriptive statements and image captures, Claude designed a website with my preferred color scheme and look. It happened right in front of me.
Here is a youtube video showing taking a napkin sketch and turning it into a web design.

MCP servers can talk to your web browser to help debug websites. Since I’m not a great graphic designer but know what I like, I describe basics, refine descriptions using an LLM, combine this with napkin sketches, and create prototypes I like.

Practical Application: Flutter Development

This approach works for difficult tasks like Flutter development. Flutter is a useful cross-platform framework but a pain to develop—all widgets must be described in Dart, a language specific to developing in Flutter. Using Figma or Penpot designs as references with an AI coding agent, creates widgets that work properly, opening doors to cross platform Android and iOS app development.

You Still Need to Understand the Fundamentals

You still must test applications because AI agents don’t necessarily make correct assumptions about your system or server. You must verify their assumptions match reality.

You still need to know how to code and set up Docker instances. You can ask AI for assistance, but there’s much AI won’t do for you—and that’s OK. It handles heavy lifting and helps with cognitive load.

Working Within Constraints

For those saying AI can’t do everything or write code right out of the box when given difficult problems—you wouldn’t do that with a junior engineer. Work with constraints. As Eli Goldratt explains in The Goal about the Theory of Constraints, you leverage limitations.

LLMs struggle with giant monolithic codebases. However, decomposing problems into smaller, modular chunks allows AI to write complex applications.

Let AI do its thing. AI handles smaller details well, though you must still test the application.

Conclusion: Stop Playing Telephone

You need good communication. As with any relationship, make everything clear and you know where you stand with the person. Establish agreements: how you’ll communicate, what norms exist, how you’ll interact with others in your group, and honor those agreements. The same applies when working with AI.

Rethink how you approach tool limitations and learn to work around constraints. Context windows, resources, and LLM abilities may someday match senior-level programmers. Meanwhile, learn to work with constraints and make communications better and concise.

Stop playing telephone with your AI, start learning how to communicate better with it and give it some guardrails.

References

Books

Tools & Technologies

Concepts