Phase Zero: How to Really Collaborate with AI

Introduction

Have you ever felt like your AI writing assistant was taking over your project? You start with a general idea, ask for some help, and suddenly the AI is generating reams of text, heading in a direction you didn’t quite intend. You’re not alone. While AI tools like ChatGPT and other Large Language Models (LLMs) can be incredibly powerful, they can also be frustratingly difficult to control. The key is learning how to collaborate effectively, rather than simply delegating.

This post will explore a better way to collaborate with AI on content creation, drawing lessons from a recent, real-world experience: the process of creating a blog post analyzing Carl Brown’s video on AGI, DeepSeek, and the future of AI. In that initial collaboration, we encountered the very problem described above – the AI “jumping the gun” with solutions before we’d fully established a shared understanding. This experience led us to develop the “Phase Zero” approach, a method for establishing a collaborative dialogue before diving into specific tasks. We’ll share that approach, and show how it can lead to a more focused and efficient workflow. You can see the blog post that inspired this improved approach.

The “Jumping the Gun” Problem: Why AI Collaboration Can Go Wrong

Large Language Models (LLMs) are designed to be helpful. They’re trained to respond to prompts quickly and comprehensively, often offering solutions before you’ve even fully articulated the problem. This eagerness can be beneficial, but it can also lead to problems:

• Premature Solutions: The LLM might generate an outline, draft text, or suggest ideas that don’t align with your overall vision.
• Incorrect Assumptions: The LLM might make assumptions about your goals, target audience, or preferred style that aren’t accurate.
• Lack of Context: Without sufficient context, the LLM’s responses might be generic, irrelevant, or even misleading.

This often results in a feeling of losing control, requiring you to spend time and effort steering the AI back on track. Our initial interaction in developing the Carl Brown blog post, in fact, exemplified this very problem. For example, you might ask an AI to write a product description, only to find it generates overly enthusiastic marketing copy that doesn’t match your brand’s voice.

Markdown: The Key to Clear Communication with AI

Before we dive into the solution, let’s talk about a simple but powerful tool that can significantly improve your interactions with AI language models: Markdown.

Markdown is a lightweight markup language that allows you to format plain text using simple symbols. Think of it as a shorthand for creating headings, lists, links, and other basic formatting elements, without the complexity of HTML or the visual clutter of a rich text editor. It’s designed to be easy to read and write, and it can be easily converted to HTML for publishing online. You don’t need to be a coder to use Markdown. It’s surprisingly simple, and even mastering just a few basic elements can make a big difference.

Why use Markdown for AI prompts?

• Clarity and Structure: Markdown’s simple syntax helps you organize your thoughts and present your instructions in a clear, structured way. This reduces ambiguity and helps the AI understand your request.
• Consistency: Using Markdown consistently ensures that all your prompts and instructions have a uniform format, making them easier to manage and reuse.
• Reduced Ambiguity: Plain text, formatted with Markdown, is less prone to misinterpretation by the AI than text copied from a rich text editor, which might contain hidden formatting codes.
• Easy Conversion to HTML: If you’re using Markdown for outlining or drafting content, it can be easily converted to HTML for publishing on your blog or website.
• Plain Text Advantage As the prompt to the LLM is text, you are sure to avoid strange formatting, and the AI can also read your markdown text.

Here are some basic Markdown syntax elements:

• Headings: Use # symbols for headings. # for H1, ## for H2, ### for H3, and so on.
• Bulleted Lists: Use * or - followed by a space to create bulleted lists.
• Links: Use [link text](URL) to create hyperlinks.
• Emphasis: Use *asterisks* for italics and **double asterisks** for bold.

For a more comprehensive guide to Markdown, check out this blogpost on Foodcourtification and this cheat sheet: Markdown Guide, Daring Fireball.

Phase Zero: Establishing a Shared Understanding with AI

The key to effective AI collaboration is establishing a shared understanding before diving into specific tasks. This is where the Phase Zero method comes in. Instead of immediately requesting a specific output (like a blog post outline), you start with a meta-prompt that sets the stage for a collaborative dialogue.

Here’s an example of a “Phase Zero” prompt:

# Project Setup: Exploring a Potential Blog Post Idea

**Goal:** I'm considering writing a blog post based on a video by Carl Brown (YouTube channel "Internet of Bugs"). I'm not yet sure of the exact focus or structure of the post. I need your help to explore the possibilities.

**Context:**

*   I'm interested in summarizing and analyzing Brown's arguments.
*   I might want to include information from other sources that Brown references.
*   The target audience is a tech-interested general public.
*   I want the blog post to be informative, engaging, and SEO-friendly.
*   I want to clarify my ideas *before* we create a detailed outline.

**Instructions:**

1.  **Ask Clarifying Questions:** Before offering any suggestions, ask me at least five questions to better understand my goals, preferences, and any existing ideas I have about the blog post. These questions should cover:
    *   The specific video I'm interested in.
    *   My initial thoughts on the main takeaways from the video.
    *   Any specific angles or themes I'm considering.
    *   My target audience (beyond the general description).
    *   Any existing content I have that might be relevant (e.g., previous blog posts).
2.  **Active Listening:** Pay close attention to my answers and use them to refine your understanding of the project.
3.  **No Premature Outlines:** Do *not* provide a blog post outline until I explicitly request it. This is a brainstorming phase.

**Desired Output (Phase Zero):** A list of clarifying questions, followed by a brief confirmation that you understand my responses.

This prompt explicitly instructs the LLM to ask questions first, forcing a “probing” approach. It emphasizes active listening and explicitly prohibits premature solutions (see below). This establishes a collaborative dialogue and ensures that the AI’s assistance is aligned with your evolving understanding of the project.

Case Study: The Carl Brown Blog Post

Our own collaboration in creating the blog post analyzing Carl Brown’s video on AGI, AGI Reality Check: DeepSeek and the Future of AI provides a real-world illustration of both the problem and the solution.

Our initial approach, while ultimately successful, highlighted a common frustration when working with AI. We began with a relatively brief prompt, asking for an outline and summaries. The AI, eager to please, immediately delivered a detailed outline. Helpful, yes – but also premature. We hadn’t yet wrestled with the core arguments, the best structure, or the ideal way to weave in those supporting sources. This led to a feeling of being behind the curve, constantly refining and redirecting the AI’s output. The “Aha!” moment – and the real impetus for this post – came with the realization that we were playing catch-up, reacting to the AI’s initiative rather than guiding it. This is the classic “jumping the gun” problem: when the tool dictates the workflow instead of the other way around.

Had we started with a Phase Zero approach, the process would likely have been smoother. The AI would have asked clarifying questions about my goals, my initial understanding of Brown’s arguments, and my preferences for integrating the supporting sources. This would have established a shared understanding before any outline was generated, leading to a more focused and efficient collaboration from the start.

Despite the initial inefficiencies, the final blog post – analyzing Carl Brown’s insights on AGI, DeepSeek, and the broader AI landscape – demonstrates the power of human-AI collaboration. The AI’s ability to quickly summarize complex information, suggest integration strategies, and provide structural guidance was invaluable. My role, as the human author, was to provide the critical judgment, editorial oversight, and contextual understanding that the AI lacked.

Practical Advice: Implementing the Phase Zero Method

Here’s how you can apply the Phase Zero method to your own AI-assisted writing projects:

1. Start with a Meta-Prompt: Before requesting any specific output, use a “Phase Zero” prompt to establish a collaborative dialogue.
2. Ask Clarifying Questions (Through the AI): Instruct the AI to ask you questions about your goals, audience, and existing ideas.
3. Define Roles Clearly: Explicitly state your role and the AI’s role in the project.
4. Iterate and Refine: Use the AI’s questions and your answers to refine your understanding of the project before moving on to specific tasks.
5. Use Markdown for Structure and Clarity: Use Markdown to format your prompts and instructions, ensuring clarity and consistency.
6. Use a template: Create a “Phase Zero” template for yourself.
7. Reflect and refine: After each project reflect on the process, what was good, what could be made even better.

The Future of Human-AI Collaboration

The experience of creating the Carl Brown blog post highlights the evolving relationship between humans and AI in content creation. AI is a powerful tool, but it’s not a replacement for human judgment, creativity, and critical thinking. The Phase Zero approach represents a shift towards a more collaborative and controlled workflow, where the human remains firmly in charge, guiding the AI’s assistance to achieve the best possible results. Mastering this AI collaboration workflow will be increasingly important. For example, imagine researchers using AI to analyze vast datasets and identify potential drug candidates, while human scientists focus on the ethical considerations and experimental design. Finding the optimal workflows will be an ongoing process of experimentation. Specifically, improving prompt engineering is crucial for interacting.

Conclusion

The Phase Zero approach offers a valuable framework for improving human-AI collaboration, avoiding the common pitfalls of premature solutions and misaligned expectations. By prioritizing shared understanding and establishing a collaborative dialogue before diving into task execution, we can harness the power of AI tools more effectively and create higher-quality content. The Carl Brown blog post serves as a concrete example of this process in action. We encourage you to try the Phase Zero approach in your own projects and share your experiences!

And, once again, here’s the link to the finished blog post we created together: AGI Reality Check: DeepSeek and the Future of AI.